[AArch64] PR target/79913: VEC_SELECT bugs in aarch64 patterns
[official-gcc.git] / gcc / calls.c
blob6d5ef4e02a01cf37b41b6d78b66950127f5f0faf
1 /* Convert function calls to rtl insns, for GNU C compiler.
2 Copyright (C) 1989-2017 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 "backend.h"
24 #include "target.h"
25 #include "rtl.h"
26 #include "tree.h"
27 #include "gimple.h"
28 #include "predict.h"
29 #include "memmodel.h"
30 #include "tm_p.h"
31 #include "stringpool.h"
32 #include "expmed.h"
33 #include "optabs.h"
34 #include "emit-rtl.h"
35 #include "cgraph.h"
36 #include "diagnostic-core.h"
37 #include "fold-const.h"
38 #include "stor-layout.h"
39 #include "varasm.h"
40 #include "internal-fn.h"
41 #include "dojump.h"
42 #include "explow.h"
43 #include "calls.h"
44 #include "expr.h"
45 #include "output.h"
46 #include "langhooks.h"
47 #include "except.h"
48 #include "dbgcnt.h"
49 #include "rtl-iter.h"
50 #include "tree-chkp.h"
51 #include "tree-vrp.h"
52 #include "tree-ssanames.h"
53 #include "rtl-chkp.h"
54 #include "intl.h"
56 /* Like PREFERRED_STACK_BOUNDARY but in units of bytes, not bits. */
57 #define STACK_BYTES (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT)
59 /* Data structure and subroutines used within expand_call. */
61 struct arg_data
63 /* Tree node for this argument. */
64 tree tree_value;
65 /* Mode for value; TYPE_MODE unless promoted. */
66 machine_mode mode;
67 /* Current RTL value for argument, or 0 if it isn't precomputed. */
68 rtx value;
69 /* Initially-compute RTL value for argument; only for const functions. */
70 rtx initial_value;
71 /* Register to pass this argument in, 0 if passed on stack, or an
72 PARALLEL if the arg is to be copied into multiple non-contiguous
73 registers. */
74 rtx reg;
75 /* Register to pass this argument in when generating tail call sequence.
76 This is not the same register as for normal calls on machines with
77 register windows. */
78 rtx tail_call_reg;
79 /* If REG is a PARALLEL, this is a copy of VALUE pulled into the correct
80 form for emit_group_move. */
81 rtx parallel_value;
82 /* If value is passed in neither reg nor stack, this field holds a number
83 of a special slot to be used. */
84 rtx special_slot;
85 /* For pointer bounds hold an index of parm bounds are bound to. -1 if
86 there is no such pointer. */
87 int pointer_arg;
88 /* If pointer_arg refers a structure, then pointer_offset holds an offset
89 of a pointer in this structure. */
90 int pointer_offset;
91 /* If REG was promoted from the actual mode of the argument expression,
92 indicates whether the promotion is sign- or zero-extended. */
93 int unsignedp;
94 /* Number of bytes to put in registers. 0 means put the whole arg
95 in registers. Also 0 if not passed in registers. */
96 int partial;
97 /* Nonzero if argument must be passed on stack.
98 Note that some arguments may be passed on the stack
99 even though pass_on_stack is zero, just because FUNCTION_ARG says so.
100 pass_on_stack identifies arguments that *cannot* go in registers. */
101 int pass_on_stack;
102 /* Some fields packaged up for locate_and_pad_parm. */
103 struct locate_and_pad_arg_data locate;
104 /* Location on the stack at which parameter should be stored. The store
105 has already been done if STACK == VALUE. */
106 rtx stack;
107 /* Location on the stack of the start of this argument slot. This can
108 differ from STACK if this arg pads downward. This location is known
109 to be aligned to TARGET_FUNCTION_ARG_BOUNDARY. */
110 rtx stack_slot;
111 /* Place that this stack area has been saved, if needed. */
112 rtx save_area;
113 /* If an argument's alignment does not permit direct copying into registers,
114 copy in smaller-sized pieces into pseudos. These are stored in a
115 block pointed to by this field. The next field says how many
116 word-sized pseudos we made. */
117 rtx *aligned_regs;
118 int n_aligned_regs;
121 /* A vector of one char per byte of stack space. A byte if nonzero if
122 the corresponding stack location has been used.
123 This vector is used to prevent a function call within an argument from
124 clobbering any stack already set up. */
125 static char *stack_usage_map;
127 /* Size of STACK_USAGE_MAP. */
128 static int highest_outgoing_arg_in_use;
130 /* A bitmap of virtual-incoming stack space. Bit is set if the corresponding
131 stack location's tail call argument has been already stored into the stack.
132 This bitmap is used to prevent sibling call optimization if function tries
133 to use parent's incoming argument slots when they have been already
134 overwritten with tail call arguments. */
135 static sbitmap stored_args_map;
137 /* stack_arg_under_construction is nonzero when an argument may be
138 initialized with a constructor call (including a C function that
139 returns a BLKmode struct) and expand_call must take special action
140 to make sure the object being constructed does not overlap the
141 argument list for the constructor call. */
142 static int stack_arg_under_construction;
144 static void emit_call_1 (rtx, tree, tree, tree, HOST_WIDE_INT, HOST_WIDE_INT,
145 HOST_WIDE_INT, rtx, rtx, int, rtx, int,
146 cumulative_args_t);
147 static void precompute_register_parameters (int, struct arg_data *, int *);
148 static void store_bounds (struct arg_data *, struct arg_data *);
149 static int store_one_arg (struct arg_data *, rtx, int, int, int);
150 static void store_unaligned_arguments_into_pseudos (struct arg_data *, int);
151 static int finalize_must_preallocate (int, int, struct arg_data *,
152 struct args_size *);
153 static void precompute_arguments (int, struct arg_data *);
154 static int compute_argument_block_size (int, struct args_size *, tree, tree, int);
155 static void initialize_argument_information (int, struct arg_data *,
156 struct args_size *, int,
157 tree, tree,
158 tree, tree, cumulative_args_t, int,
159 rtx *, int *, int *, int *,
160 bool *, bool);
161 static void compute_argument_addresses (struct arg_data *, rtx, int);
162 static rtx rtx_for_function_call (tree, tree);
163 static void load_register_parameters (struct arg_data *, int, rtx *, int,
164 int, int *);
165 static rtx emit_library_call_value_1 (int, rtx, rtx, enum libcall_type,
166 machine_mode, int, va_list);
167 static int special_function_p (const_tree, int);
168 static int check_sibcall_argument_overlap_1 (rtx);
169 static int check_sibcall_argument_overlap (rtx_insn *, struct arg_data *, int);
171 static int combine_pending_stack_adjustment_and_call (int, struct args_size *,
172 unsigned int);
173 static tree split_complex_types (tree);
175 #ifdef REG_PARM_STACK_SPACE
176 static rtx save_fixed_argument_area (int, rtx, int *, int *);
177 static void restore_fixed_argument_area (rtx, rtx, int, int);
178 #endif
180 /* Force FUNEXP into a form suitable for the address of a CALL,
181 and return that as an rtx. Also load the static chain register
182 if FNDECL is a nested function.
184 CALL_FUSAGE points to a variable holding the prospective
185 CALL_INSN_FUNCTION_USAGE information. */
188 prepare_call_address (tree fndecl_or_type, rtx funexp, rtx static_chain_value,
189 rtx *call_fusage, int reg_parm_seen, int flags)
191 /* Make a valid memory address and copy constants through pseudo-regs,
192 but not for a constant address if -fno-function-cse. */
193 if (GET_CODE (funexp) != SYMBOL_REF)
195 /* If it's an indirect call by descriptor, generate code to perform
196 runtime identification of the pointer and load the descriptor. */
197 if ((flags & ECF_BY_DESCRIPTOR) && !flag_trampolines)
199 const int bit_val = targetm.calls.custom_function_descriptors;
200 rtx call_lab = gen_label_rtx ();
202 gcc_assert (fndecl_or_type && TYPE_P (fndecl_or_type));
203 fndecl_or_type
204 = build_decl (UNKNOWN_LOCATION, FUNCTION_DECL, NULL_TREE,
205 fndecl_or_type);
206 DECL_STATIC_CHAIN (fndecl_or_type) = 1;
207 rtx chain = targetm.calls.static_chain (fndecl_or_type, false);
209 /* Avoid long live ranges around function calls. */
210 funexp = copy_to_mode_reg (Pmode, funexp);
212 if (REG_P (chain))
213 emit_insn (gen_rtx_CLOBBER (VOIDmode, chain));
215 /* Emit the runtime identification pattern. */
216 rtx mask = gen_rtx_AND (Pmode, funexp, GEN_INT (bit_val));
217 emit_cmp_and_jump_insns (mask, const0_rtx, EQ, NULL_RTX, Pmode, 1,
218 call_lab);
220 /* Statically predict the branch to very likely taken. */
221 rtx_insn *insn = get_last_insn ();
222 if (JUMP_P (insn))
223 predict_insn_def (insn, PRED_BUILTIN_EXPECT, TAKEN);
225 /* Load the descriptor. */
226 rtx mem = gen_rtx_MEM (ptr_mode,
227 plus_constant (Pmode, funexp, - bit_val));
228 MEM_NOTRAP_P (mem) = 1;
229 mem = convert_memory_address (Pmode, mem);
230 emit_move_insn (chain, mem);
232 mem = gen_rtx_MEM (ptr_mode,
233 plus_constant (Pmode, funexp,
234 POINTER_SIZE / BITS_PER_UNIT
235 - bit_val));
236 MEM_NOTRAP_P (mem) = 1;
237 mem = convert_memory_address (Pmode, mem);
238 emit_move_insn (funexp, mem);
240 emit_label (call_lab);
242 if (REG_P (chain))
244 use_reg (call_fusage, chain);
245 STATIC_CHAIN_REG_P (chain) = 1;
248 /* Make sure we're not going to be overwritten below. */
249 gcc_assert (!static_chain_value);
252 /* If we are using registers for parameters, force the
253 function address into a register now. */
254 funexp = ((reg_parm_seen
255 && targetm.small_register_classes_for_mode_p (FUNCTION_MODE))
256 ? force_not_mem (memory_address (FUNCTION_MODE, funexp))
257 : memory_address (FUNCTION_MODE, funexp));
259 else
261 /* funexp could be a SYMBOL_REF represents a function pointer which is
262 of ptr_mode. In this case, it should be converted into address mode
263 to be a valid address for memory rtx pattern. See PR 64971. */
264 if (GET_MODE (funexp) != Pmode)
265 funexp = convert_memory_address (Pmode, funexp);
267 if (!(flags & ECF_SIBCALL))
269 if (!NO_FUNCTION_CSE && optimize && ! flag_no_function_cse)
270 funexp = force_reg (Pmode, funexp);
274 if (static_chain_value != 0
275 && (TREE_CODE (fndecl_or_type) != FUNCTION_DECL
276 || DECL_STATIC_CHAIN (fndecl_or_type)))
278 rtx chain;
280 chain = targetm.calls.static_chain (fndecl_or_type, false);
281 static_chain_value = convert_memory_address (Pmode, static_chain_value);
283 emit_move_insn (chain, static_chain_value);
284 if (REG_P (chain))
286 use_reg (call_fusage, chain);
287 STATIC_CHAIN_REG_P (chain) = 1;
291 return funexp;
294 /* Generate instructions to call function FUNEXP,
295 and optionally pop the results.
296 The CALL_INSN is the first insn generated.
298 FNDECL is the declaration node of the function. This is given to the
299 hook TARGET_RETURN_POPS_ARGS to determine whether this function pops
300 its own args.
302 FUNTYPE is the data type of the function. This is given to the hook
303 TARGET_RETURN_POPS_ARGS to determine whether this function pops its
304 own args. We used to allow an identifier for library functions, but
305 that doesn't work when the return type is an aggregate type and the
306 calling convention says that the pointer to this aggregate is to be
307 popped by the callee.
309 STACK_SIZE is the number of bytes of arguments on the stack,
310 ROUNDED_STACK_SIZE is that number rounded up to
311 PREFERRED_STACK_BOUNDARY; zero if the size is variable. This is
312 both to put into the call insn and to generate explicit popping
313 code if necessary.
315 STRUCT_VALUE_SIZE is the number of bytes wanted in a structure value.
316 It is zero if this call doesn't want a structure value.
318 NEXT_ARG_REG is the rtx that results from executing
319 targetm.calls.function_arg (&args_so_far, VOIDmode, void_type_node, true)
320 just after all the args have had their registers assigned.
321 This could be whatever you like, but normally it is the first
322 arg-register beyond those used for args in this call,
323 or 0 if all the arg-registers are used in this call.
324 It is passed on to `gen_call' so you can put this info in the call insn.
326 VALREG is a hard register in which a value is returned,
327 or 0 if the call does not return a value.
329 OLD_INHIBIT_DEFER_POP is the value that `inhibit_defer_pop' had before
330 the args to this call were processed.
331 We restore `inhibit_defer_pop' to that value.
333 CALL_FUSAGE is either empty or an EXPR_LIST of USE expressions that
334 denote registers used by the called function. */
336 static void
337 emit_call_1 (rtx funexp, tree fntree ATTRIBUTE_UNUSED, tree fndecl ATTRIBUTE_UNUSED,
338 tree funtype ATTRIBUTE_UNUSED,
339 HOST_WIDE_INT stack_size ATTRIBUTE_UNUSED,
340 HOST_WIDE_INT rounded_stack_size,
341 HOST_WIDE_INT struct_value_size ATTRIBUTE_UNUSED,
342 rtx next_arg_reg ATTRIBUTE_UNUSED, rtx valreg,
343 int old_inhibit_defer_pop, rtx call_fusage, int ecf_flags,
344 cumulative_args_t args_so_far ATTRIBUTE_UNUSED)
346 rtx rounded_stack_size_rtx = GEN_INT (rounded_stack_size);
347 rtx call, funmem, pat;
348 int already_popped = 0;
349 HOST_WIDE_INT n_popped = 0;
351 /* Sibling call patterns never pop arguments (no sibcall(_value)_pop
352 patterns exist). Any popping that the callee does on return will
353 be from our caller's frame rather than ours. */
354 if (!(ecf_flags & ECF_SIBCALL))
356 n_popped += targetm.calls.return_pops_args (fndecl, funtype, stack_size);
358 #ifdef CALL_POPS_ARGS
359 n_popped += CALL_POPS_ARGS (*get_cumulative_args (args_so_far));
360 #endif
363 /* Ensure address is valid. SYMBOL_REF is already valid, so no need,
364 and we don't want to load it into a register as an optimization,
365 because prepare_call_address already did it if it should be done. */
366 if (GET_CODE (funexp) != SYMBOL_REF)
367 funexp = memory_address (FUNCTION_MODE, funexp);
369 funmem = gen_rtx_MEM (FUNCTION_MODE, funexp);
370 if (fndecl && TREE_CODE (fndecl) == FUNCTION_DECL)
372 tree t = fndecl;
374 /* Although a built-in FUNCTION_DECL and its non-__builtin
375 counterpart compare equal and get a shared mem_attrs, they
376 produce different dump output in compare-debug compilations,
377 if an entry gets garbage collected in one compilation, then
378 adds a different (but equivalent) entry, while the other
379 doesn't run the garbage collector at the same spot and then
380 shares the mem_attr with the equivalent entry. */
381 if (DECL_BUILT_IN_CLASS (t) == BUILT_IN_NORMAL)
383 tree t2 = builtin_decl_explicit (DECL_FUNCTION_CODE (t));
384 if (t2)
385 t = t2;
388 set_mem_expr (funmem, t);
390 else if (fntree)
391 set_mem_expr (funmem, build_simple_mem_ref (CALL_EXPR_FN (fntree)));
393 if (ecf_flags & ECF_SIBCALL)
395 if (valreg)
396 pat = targetm.gen_sibcall_value (valreg, funmem,
397 rounded_stack_size_rtx,
398 next_arg_reg, NULL_RTX);
399 else
400 pat = targetm.gen_sibcall (funmem, rounded_stack_size_rtx,
401 next_arg_reg, GEN_INT (struct_value_size));
403 /* If the target has "call" or "call_value" insns, then prefer them
404 if no arguments are actually popped. If the target does not have
405 "call" or "call_value" insns, then we must use the popping versions
406 even if the call has no arguments to pop. */
407 else if (n_popped > 0
408 || !(valreg
409 ? targetm.have_call_value ()
410 : targetm.have_call ()))
412 rtx n_pop = GEN_INT (n_popped);
414 /* If this subroutine pops its own args, record that in the call insn
415 if possible, for the sake of frame pointer elimination. */
417 if (valreg)
418 pat = targetm.gen_call_value_pop (valreg, funmem,
419 rounded_stack_size_rtx,
420 next_arg_reg, n_pop);
421 else
422 pat = targetm.gen_call_pop (funmem, rounded_stack_size_rtx,
423 next_arg_reg, n_pop);
425 already_popped = 1;
427 else
429 if (valreg)
430 pat = targetm.gen_call_value (valreg, funmem, rounded_stack_size_rtx,
431 next_arg_reg, NULL_RTX);
432 else
433 pat = targetm.gen_call (funmem, rounded_stack_size_rtx, next_arg_reg,
434 GEN_INT (struct_value_size));
436 emit_insn (pat);
438 /* Find the call we just emitted. */
439 rtx_call_insn *call_insn = last_call_insn ();
441 /* Some target create a fresh MEM instead of reusing the one provided
442 above. Set its MEM_EXPR. */
443 call = get_call_rtx_from (call_insn);
444 if (call
445 && MEM_EXPR (XEXP (call, 0)) == NULL_TREE
446 && MEM_EXPR (funmem) != NULL_TREE)
447 set_mem_expr (XEXP (call, 0), MEM_EXPR (funmem));
449 /* Mark instrumented calls. */
450 if (call && fntree)
451 CALL_EXPR_WITH_BOUNDS_P (call) = CALL_WITH_BOUNDS_P (fntree);
453 /* Put the register usage information there. */
454 add_function_usage_to (call_insn, call_fusage);
456 /* If this is a const call, then set the insn's unchanging bit. */
457 if (ecf_flags & ECF_CONST)
458 RTL_CONST_CALL_P (call_insn) = 1;
460 /* If this is a pure call, then set the insn's unchanging bit. */
461 if (ecf_flags & ECF_PURE)
462 RTL_PURE_CALL_P (call_insn) = 1;
464 /* If this is a const call, then set the insn's unchanging bit. */
465 if (ecf_flags & ECF_LOOPING_CONST_OR_PURE)
466 RTL_LOOPING_CONST_OR_PURE_CALL_P (call_insn) = 1;
468 /* Create a nothrow REG_EH_REGION note, if needed. */
469 make_reg_eh_region_note (call_insn, ecf_flags, 0);
471 if (ecf_flags & ECF_NORETURN)
472 add_reg_note (call_insn, REG_NORETURN, const0_rtx);
474 if (ecf_flags & ECF_RETURNS_TWICE)
476 add_reg_note (call_insn, REG_SETJMP, const0_rtx);
477 cfun->calls_setjmp = 1;
480 SIBLING_CALL_P (call_insn) = ((ecf_flags & ECF_SIBCALL) != 0);
482 /* Restore this now, so that we do defer pops for this call's args
483 if the context of the call as a whole permits. */
484 inhibit_defer_pop = old_inhibit_defer_pop;
486 if (n_popped > 0)
488 if (!already_popped)
489 CALL_INSN_FUNCTION_USAGE (call_insn)
490 = gen_rtx_EXPR_LIST (VOIDmode,
491 gen_rtx_CLOBBER (VOIDmode, stack_pointer_rtx),
492 CALL_INSN_FUNCTION_USAGE (call_insn));
493 rounded_stack_size -= n_popped;
494 rounded_stack_size_rtx = GEN_INT (rounded_stack_size);
495 stack_pointer_delta -= n_popped;
497 add_reg_note (call_insn, REG_ARGS_SIZE, GEN_INT (stack_pointer_delta));
499 /* If popup is needed, stack realign must use DRAP */
500 if (SUPPORTS_STACK_ALIGNMENT)
501 crtl->need_drap = true;
503 /* For noreturn calls when not accumulating outgoing args force
504 REG_ARGS_SIZE note to prevent crossjumping of calls with different
505 args sizes. */
506 else if (!ACCUMULATE_OUTGOING_ARGS && (ecf_flags & ECF_NORETURN) != 0)
507 add_reg_note (call_insn, REG_ARGS_SIZE, GEN_INT (stack_pointer_delta));
509 if (!ACCUMULATE_OUTGOING_ARGS)
511 /* If returning from the subroutine does not automatically pop the args,
512 we need an instruction to pop them sooner or later.
513 Perhaps do it now; perhaps just record how much space to pop later.
515 If returning from the subroutine does pop the args, indicate that the
516 stack pointer will be changed. */
518 if (rounded_stack_size != 0)
520 if (ecf_flags & ECF_NORETURN)
521 /* Just pretend we did the pop. */
522 stack_pointer_delta -= rounded_stack_size;
523 else if (flag_defer_pop && inhibit_defer_pop == 0
524 && ! (ecf_flags & (ECF_CONST | ECF_PURE)))
525 pending_stack_adjust += rounded_stack_size;
526 else
527 adjust_stack (rounded_stack_size_rtx);
530 /* When we accumulate outgoing args, we must avoid any stack manipulations.
531 Restore the stack pointer to its original value now. Usually
532 ACCUMULATE_OUTGOING_ARGS targets don't get here, but there are exceptions.
533 On i386 ACCUMULATE_OUTGOING_ARGS can be enabled on demand, and
534 popping variants of functions exist as well.
536 ??? We may optimize similar to defer_pop above, but it is
537 probably not worthwhile.
539 ??? It will be worthwhile to enable combine_stack_adjustments even for
540 such machines. */
541 else if (n_popped)
542 anti_adjust_stack (GEN_INT (n_popped));
545 /* Determine if the function identified by FNDECL is one with
546 special properties we wish to know about. Modify FLAGS accordingly.
548 For example, if the function might return more than one time (setjmp), then
549 set ECF_RETURNS_TWICE.
551 Set ECF_MAY_BE_ALLOCA for any memory allocation function that might allocate
552 space from the stack such as alloca. */
554 static int
555 special_function_p (const_tree fndecl, int flags)
557 tree name_decl = DECL_NAME (fndecl);
559 /* For instrumentation clones we want to derive flags
560 from the original name. */
561 if (cgraph_node::get (fndecl)
562 && cgraph_node::get (fndecl)->instrumentation_clone)
563 name_decl = DECL_NAME (cgraph_node::get (fndecl)->orig_decl);
565 if (fndecl && name_decl
566 && IDENTIFIER_LENGTH (name_decl) <= 11
567 /* Exclude functions not at the file scope, or not `extern',
568 since they are not the magic functions we would otherwise
569 think they are.
570 FIXME: this should be handled with attributes, not with this
571 hacky imitation of DECL_ASSEMBLER_NAME. It's (also) wrong
572 because you can declare fork() inside a function if you
573 wish. */
574 && (DECL_CONTEXT (fndecl) == NULL_TREE
575 || TREE_CODE (DECL_CONTEXT (fndecl)) == TRANSLATION_UNIT_DECL)
576 && TREE_PUBLIC (fndecl))
578 const char *name = IDENTIFIER_POINTER (name_decl);
579 const char *tname = name;
581 /* We assume that alloca will always be called by name. It
582 makes no sense to pass it as a pointer-to-function to
583 anything that does not understand its behavior. */
584 if (IDENTIFIER_LENGTH (name_decl) == 6
585 && name[0] == 'a'
586 && ! strcmp (name, "alloca"))
587 flags |= ECF_MAY_BE_ALLOCA;
589 /* Disregard prefix _ or __. */
590 if (name[0] == '_')
592 if (name[1] == '_')
593 tname += 2;
594 else
595 tname += 1;
598 /* ECF_RETURNS_TWICE is safe even for -ffreestanding. */
599 if (! strcmp (tname, "setjmp")
600 || ! strcmp (tname, "sigsetjmp")
601 || ! strcmp (name, "savectx")
602 || ! strcmp (name, "vfork")
603 || ! strcmp (name, "getcontext"))
604 flags |= ECF_RETURNS_TWICE;
607 if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
608 switch (DECL_FUNCTION_CODE (fndecl))
610 case BUILT_IN_ALLOCA:
611 case BUILT_IN_ALLOCA_WITH_ALIGN:
612 flags |= ECF_MAY_BE_ALLOCA;
613 break;
614 default:
615 break;
618 return flags;
621 /* Similar to special_function_p; return a set of ERF_ flags for the
622 function FNDECL. */
623 static int
624 decl_return_flags (tree fndecl)
626 tree attr;
627 tree type = TREE_TYPE (fndecl);
628 if (!type)
629 return 0;
631 attr = lookup_attribute ("fn spec", TYPE_ATTRIBUTES (type));
632 if (!attr)
633 return 0;
635 attr = TREE_VALUE (TREE_VALUE (attr));
636 if (!attr || TREE_STRING_LENGTH (attr) < 1)
637 return 0;
639 switch (TREE_STRING_POINTER (attr)[0])
641 case '1':
642 case '2':
643 case '3':
644 case '4':
645 return ERF_RETURNS_ARG | (TREE_STRING_POINTER (attr)[0] - '1');
647 case 'm':
648 return ERF_NOALIAS;
650 case '.':
651 default:
652 return 0;
656 /* Return nonzero when FNDECL represents a call to setjmp. */
659 setjmp_call_p (const_tree fndecl)
661 if (DECL_IS_RETURNS_TWICE (fndecl))
662 return ECF_RETURNS_TWICE;
663 return special_function_p (fndecl, 0) & ECF_RETURNS_TWICE;
667 /* Return true if STMT may be an alloca call. */
669 bool
670 gimple_maybe_alloca_call_p (const gimple *stmt)
672 tree fndecl;
674 if (!is_gimple_call (stmt))
675 return false;
677 fndecl = gimple_call_fndecl (stmt);
678 if (fndecl && (special_function_p (fndecl, 0) & ECF_MAY_BE_ALLOCA))
679 return true;
681 return false;
684 /* Return true if STMT is a builtin alloca call. */
686 bool
687 gimple_alloca_call_p (const gimple *stmt)
689 tree fndecl;
691 if (!is_gimple_call (stmt))
692 return false;
694 fndecl = gimple_call_fndecl (stmt);
695 if (fndecl && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
696 switch (DECL_FUNCTION_CODE (fndecl))
698 case BUILT_IN_ALLOCA:
699 case BUILT_IN_ALLOCA_WITH_ALIGN:
700 return true;
701 default:
702 break;
705 return false;
708 /* Return true when exp contains a builtin alloca call. */
710 bool
711 alloca_call_p (const_tree exp)
713 tree fndecl;
714 if (TREE_CODE (exp) == CALL_EXPR
715 && (fndecl = get_callee_fndecl (exp))
716 && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
717 switch (DECL_FUNCTION_CODE (fndecl))
719 case BUILT_IN_ALLOCA:
720 case BUILT_IN_ALLOCA_WITH_ALIGN:
721 return true;
722 default:
723 break;
726 return false;
729 /* Return TRUE if FNDECL is either a TM builtin or a TM cloned
730 function. Return FALSE otherwise. */
732 static bool
733 is_tm_builtin (const_tree fndecl)
735 if (fndecl == NULL)
736 return false;
738 if (decl_is_tm_clone (fndecl))
739 return true;
741 if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
743 switch (DECL_FUNCTION_CODE (fndecl))
745 case BUILT_IN_TM_COMMIT:
746 case BUILT_IN_TM_COMMIT_EH:
747 case BUILT_IN_TM_ABORT:
748 case BUILT_IN_TM_IRREVOCABLE:
749 case BUILT_IN_TM_GETTMCLONE_IRR:
750 case BUILT_IN_TM_MEMCPY:
751 case BUILT_IN_TM_MEMMOVE:
752 case BUILT_IN_TM_MEMSET:
753 CASE_BUILT_IN_TM_STORE (1):
754 CASE_BUILT_IN_TM_STORE (2):
755 CASE_BUILT_IN_TM_STORE (4):
756 CASE_BUILT_IN_TM_STORE (8):
757 CASE_BUILT_IN_TM_STORE (FLOAT):
758 CASE_BUILT_IN_TM_STORE (DOUBLE):
759 CASE_BUILT_IN_TM_STORE (LDOUBLE):
760 CASE_BUILT_IN_TM_STORE (M64):
761 CASE_BUILT_IN_TM_STORE (M128):
762 CASE_BUILT_IN_TM_STORE (M256):
763 CASE_BUILT_IN_TM_LOAD (1):
764 CASE_BUILT_IN_TM_LOAD (2):
765 CASE_BUILT_IN_TM_LOAD (4):
766 CASE_BUILT_IN_TM_LOAD (8):
767 CASE_BUILT_IN_TM_LOAD (FLOAT):
768 CASE_BUILT_IN_TM_LOAD (DOUBLE):
769 CASE_BUILT_IN_TM_LOAD (LDOUBLE):
770 CASE_BUILT_IN_TM_LOAD (M64):
771 CASE_BUILT_IN_TM_LOAD (M128):
772 CASE_BUILT_IN_TM_LOAD (M256):
773 case BUILT_IN_TM_LOG:
774 case BUILT_IN_TM_LOG_1:
775 case BUILT_IN_TM_LOG_2:
776 case BUILT_IN_TM_LOG_4:
777 case BUILT_IN_TM_LOG_8:
778 case BUILT_IN_TM_LOG_FLOAT:
779 case BUILT_IN_TM_LOG_DOUBLE:
780 case BUILT_IN_TM_LOG_LDOUBLE:
781 case BUILT_IN_TM_LOG_M64:
782 case BUILT_IN_TM_LOG_M128:
783 case BUILT_IN_TM_LOG_M256:
784 return true;
785 default:
786 break;
789 return false;
792 /* Detect flags (function attributes) from the function decl or type node. */
795 flags_from_decl_or_type (const_tree exp)
797 int flags = 0;
799 if (DECL_P (exp))
801 /* The function exp may have the `malloc' attribute. */
802 if (DECL_IS_MALLOC (exp))
803 flags |= ECF_MALLOC;
805 /* The function exp may have the `returns_twice' attribute. */
806 if (DECL_IS_RETURNS_TWICE (exp))
807 flags |= ECF_RETURNS_TWICE;
809 /* Process the pure and const attributes. */
810 if (TREE_READONLY (exp))
811 flags |= ECF_CONST;
812 if (DECL_PURE_P (exp))
813 flags |= ECF_PURE;
814 if (DECL_LOOPING_CONST_OR_PURE_P (exp))
815 flags |= ECF_LOOPING_CONST_OR_PURE;
817 if (DECL_IS_NOVOPS (exp))
818 flags |= ECF_NOVOPS;
819 if (lookup_attribute ("leaf", DECL_ATTRIBUTES (exp)))
820 flags |= ECF_LEAF;
822 if (TREE_NOTHROW (exp))
823 flags |= ECF_NOTHROW;
825 if (flag_tm)
827 if (is_tm_builtin (exp))
828 flags |= ECF_TM_BUILTIN;
829 else if ((flags & (ECF_CONST|ECF_NOVOPS)) != 0
830 || lookup_attribute ("transaction_pure",
831 TYPE_ATTRIBUTES (TREE_TYPE (exp))))
832 flags |= ECF_TM_PURE;
835 flags = special_function_p (exp, flags);
837 else if (TYPE_P (exp))
839 if (TYPE_READONLY (exp))
840 flags |= ECF_CONST;
842 if (flag_tm
843 && ((flags & ECF_CONST) != 0
844 || lookup_attribute ("transaction_pure", TYPE_ATTRIBUTES (exp))))
845 flags |= ECF_TM_PURE;
847 else
848 gcc_unreachable ();
850 if (TREE_THIS_VOLATILE (exp))
852 flags |= ECF_NORETURN;
853 if (flags & (ECF_CONST|ECF_PURE))
854 flags |= ECF_LOOPING_CONST_OR_PURE;
857 return flags;
860 /* Detect flags from a CALL_EXPR. */
863 call_expr_flags (const_tree t)
865 int flags;
866 tree decl = get_callee_fndecl (t);
868 if (decl)
869 flags = flags_from_decl_or_type (decl);
870 else if (CALL_EXPR_FN (t) == NULL_TREE)
871 flags = internal_fn_flags (CALL_EXPR_IFN (t));
872 else
874 tree type = TREE_TYPE (CALL_EXPR_FN (t));
875 if (type && TREE_CODE (type) == POINTER_TYPE)
876 flags = flags_from_decl_or_type (TREE_TYPE (type));
877 else
878 flags = 0;
879 if (CALL_EXPR_BY_DESCRIPTOR (t))
880 flags |= ECF_BY_DESCRIPTOR;
883 return flags;
886 /* Return true if TYPE should be passed by invisible reference. */
888 bool
889 pass_by_reference (CUMULATIVE_ARGS *ca, machine_mode mode,
890 tree type, bool named_arg)
892 if (type)
894 /* If this type contains non-trivial constructors, then it is
895 forbidden for the middle-end to create any new copies. */
896 if (TREE_ADDRESSABLE (type))
897 return true;
899 /* GCC post 3.4 passes *all* variable sized types by reference. */
900 if (!TYPE_SIZE (type) || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
901 return true;
903 /* If a record type should be passed the same as its first (and only)
904 member, use the type and mode of that member. */
905 if (TREE_CODE (type) == RECORD_TYPE && TYPE_TRANSPARENT_AGGR (type))
907 type = TREE_TYPE (first_field (type));
908 mode = TYPE_MODE (type);
912 return targetm.calls.pass_by_reference (pack_cumulative_args (ca), mode,
913 type, named_arg);
916 /* Return true if TYPE, which is passed by reference, should be callee
917 copied instead of caller copied. */
919 bool
920 reference_callee_copied (CUMULATIVE_ARGS *ca, machine_mode mode,
921 tree type, bool named_arg)
923 if (type && TREE_ADDRESSABLE (type))
924 return false;
925 return targetm.calls.callee_copies (pack_cumulative_args (ca), mode, type,
926 named_arg);
930 /* Precompute all register parameters as described by ARGS, storing values
931 into fields within the ARGS array.
933 NUM_ACTUALS indicates the total number elements in the ARGS array.
935 Set REG_PARM_SEEN if we encounter a register parameter. */
937 static void
938 precompute_register_parameters (int num_actuals, struct arg_data *args,
939 int *reg_parm_seen)
941 int i;
943 *reg_parm_seen = 0;
945 for (i = 0; i < num_actuals; i++)
946 if (args[i].reg != 0 && ! args[i].pass_on_stack)
948 *reg_parm_seen = 1;
950 if (args[i].value == 0)
952 push_temp_slots ();
953 args[i].value = expand_normal (args[i].tree_value);
954 preserve_temp_slots (args[i].value);
955 pop_temp_slots ();
958 /* If we are to promote the function arg to a wider mode,
959 do it now. */
961 if (args[i].mode != TYPE_MODE (TREE_TYPE (args[i].tree_value)))
962 args[i].value
963 = convert_modes (args[i].mode,
964 TYPE_MODE (TREE_TYPE (args[i].tree_value)),
965 args[i].value, args[i].unsignedp);
967 /* If the value is a non-legitimate constant, force it into a
968 pseudo now. TLS symbols sometimes need a call to resolve. */
969 if (CONSTANT_P (args[i].value)
970 && !targetm.legitimate_constant_p (args[i].mode, args[i].value))
971 args[i].value = force_reg (args[i].mode, args[i].value);
973 /* If we're going to have to load the value by parts, pull the
974 parts into pseudos. The part extraction process can involve
975 non-trivial computation. */
976 if (GET_CODE (args[i].reg) == PARALLEL)
978 tree type = TREE_TYPE (args[i].tree_value);
979 args[i].parallel_value
980 = emit_group_load_into_temps (args[i].reg, args[i].value,
981 type, int_size_in_bytes (type));
984 /* If the value is expensive, and we are inside an appropriately
985 short loop, put the value into a pseudo and then put the pseudo
986 into the hard reg.
988 For small register classes, also do this if this call uses
989 register parameters. This is to avoid reload conflicts while
990 loading the parameters registers. */
992 else if ((! (REG_P (args[i].value)
993 || (GET_CODE (args[i].value) == SUBREG
994 && REG_P (SUBREG_REG (args[i].value)))))
995 && args[i].mode != BLKmode
996 && (set_src_cost (args[i].value, args[i].mode,
997 optimize_insn_for_speed_p ())
998 > COSTS_N_INSNS (1))
999 && ((*reg_parm_seen
1000 && targetm.small_register_classes_for_mode_p (args[i].mode))
1001 || optimize))
1002 args[i].value = copy_to_mode_reg (args[i].mode, args[i].value);
1006 #ifdef REG_PARM_STACK_SPACE
1008 /* The argument list is the property of the called routine and it
1009 may clobber it. If the fixed area has been used for previous
1010 parameters, we must save and restore it. */
1012 static rtx
1013 save_fixed_argument_area (int reg_parm_stack_space, rtx argblock, int *low_to_save, int *high_to_save)
1015 int low;
1016 int high;
1018 /* Compute the boundary of the area that needs to be saved, if any. */
1019 high = reg_parm_stack_space;
1020 if (ARGS_GROW_DOWNWARD)
1021 high += 1;
1023 if (high > highest_outgoing_arg_in_use)
1024 high = highest_outgoing_arg_in_use;
1026 for (low = 0; low < high; low++)
1027 if (stack_usage_map[low] != 0)
1029 int num_to_save;
1030 machine_mode save_mode;
1031 int delta;
1032 rtx addr;
1033 rtx stack_area;
1034 rtx save_area;
1036 while (stack_usage_map[--high] == 0)
1039 *low_to_save = low;
1040 *high_to_save = high;
1042 num_to_save = high - low + 1;
1043 save_mode = mode_for_size (num_to_save * BITS_PER_UNIT, MODE_INT, 1);
1045 /* If we don't have the required alignment, must do this
1046 in BLKmode. */
1047 if ((low & (MIN (GET_MODE_SIZE (save_mode),
1048 BIGGEST_ALIGNMENT / UNITS_PER_WORD) - 1)))
1049 save_mode = BLKmode;
1051 if (ARGS_GROW_DOWNWARD)
1052 delta = -high;
1053 else
1054 delta = low;
1056 addr = plus_constant (Pmode, argblock, delta);
1057 stack_area = gen_rtx_MEM (save_mode, memory_address (save_mode, addr));
1059 set_mem_align (stack_area, PARM_BOUNDARY);
1060 if (save_mode == BLKmode)
1062 save_area = assign_stack_temp (BLKmode, num_to_save);
1063 emit_block_move (validize_mem (save_area), stack_area,
1064 GEN_INT (num_to_save), BLOCK_OP_CALL_PARM);
1066 else
1068 save_area = gen_reg_rtx (save_mode);
1069 emit_move_insn (save_area, stack_area);
1072 return save_area;
1075 return NULL_RTX;
1078 static void
1079 restore_fixed_argument_area (rtx save_area, rtx argblock, int high_to_save, int low_to_save)
1081 machine_mode save_mode = GET_MODE (save_area);
1082 int delta;
1083 rtx addr, stack_area;
1085 if (ARGS_GROW_DOWNWARD)
1086 delta = -high_to_save;
1087 else
1088 delta = low_to_save;
1090 addr = plus_constant (Pmode, argblock, delta);
1091 stack_area = gen_rtx_MEM (save_mode, memory_address (save_mode, addr));
1092 set_mem_align (stack_area, PARM_BOUNDARY);
1094 if (save_mode != BLKmode)
1095 emit_move_insn (stack_area, save_area);
1096 else
1097 emit_block_move (stack_area, validize_mem (save_area),
1098 GEN_INT (high_to_save - low_to_save + 1),
1099 BLOCK_OP_CALL_PARM);
1101 #endif /* REG_PARM_STACK_SPACE */
1103 /* If any elements in ARGS refer to parameters that are to be passed in
1104 registers, but not in memory, and whose alignment does not permit a
1105 direct copy into registers. Copy the values into a group of pseudos
1106 which we will later copy into the appropriate hard registers.
1108 Pseudos for each unaligned argument will be stored into the array
1109 args[argnum].aligned_regs. The caller is responsible for deallocating
1110 the aligned_regs array if it is nonzero. */
1112 static void
1113 store_unaligned_arguments_into_pseudos (struct arg_data *args, int num_actuals)
1115 int i, j;
1117 for (i = 0; i < num_actuals; i++)
1118 if (args[i].reg != 0 && ! args[i].pass_on_stack
1119 && GET_CODE (args[i].reg) != PARALLEL
1120 && args[i].mode == BLKmode
1121 && MEM_P (args[i].value)
1122 && (MEM_ALIGN (args[i].value)
1123 < (unsigned int) MIN (BIGGEST_ALIGNMENT, BITS_PER_WORD)))
1125 int bytes = int_size_in_bytes (TREE_TYPE (args[i].tree_value));
1126 int endian_correction = 0;
1128 if (args[i].partial)
1130 gcc_assert (args[i].partial % UNITS_PER_WORD == 0);
1131 args[i].n_aligned_regs = args[i].partial / UNITS_PER_WORD;
1133 else
1135 args[i].n_aligned_regs
1136 = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
1139 args[i].aligned_regs = XNEWVEC (rtx, args[i].n_aligned_regs);
1141 /* Structures smaller than a word are normally aligned to the
1142 least significant byte. On a BYTES_BIG_ENDIAN machine,
1143 this means we must skip the empty high order bytes when
1144 calculating the bit offset. */
1145 if (bytes < UNITS_PER_WORD
1146 #ifdef BLOCK_REG_PADDING
1147 && (BLOCK_REG_PADDING (args[i].mode,
1148 TREE_TYPE (args[i].tree_value), 1)
1149 == downward)
1150 #else
1151 && BYTES_BIG_ENDIAN
1152 #endif
1154 endian_correction = BITS_PER_WORD - bytes * BITS_PER_UNIT;
1156 for (j = 0; j < args[i].n_aligned_regs; j++)
1158 rtx reg = gen_reg_rtx (word_mode);
1159 rtx word = operand_subword_force (args[i].value, j, BLKmode);
1160 int bitsize = MIN (bytes * BITS_PER_UNIT, BITS_PER_WORD);
1162 args[i].aligned_regs[j] = reg;
1163 word = extract_bit_field (word, bitsize, 0, 1, NULL_RTX,
1164 word_mode, word_mode, false);
1166 /* There is no need to restrict this code to loading items
1167 in TYPE_ALIGN sized hunks. The bitfield instructions can
1168 load up entire word sized registers efficiently.
1170 ??? This may not be needed anymore.
1171 We use to emit a clobber here but that doesn't let later
1172 passes optimize the instructions we emit. By storing 0 into
1173 the register later passes know the first AND to zero out the
1174 bitfield being set in the register is unnecessary. The store
1175 of 0 will be deleted as will at least the first AND. */
1177 emit_move_insn (reg, const0_rtx);
1179 bytes -= bitsize / BITS_PER_UNIT;
1180 store_bit_field (reg, bitsize, endian_correction, 0, 0,
1181 word_mode, word, false);
1186 /* The limit set by -Walloc-larger-than=. */
1187 static GTY(()) tree alloc_object_size_limit;
1189 /* Initialize ALLOC_OBJECT_SIZE_LIMIT based on the -Walloc-size-larger-than=
1190 setting if the option is specified, or to the maximum object size if it
1191 is not. Return the initialized value. */
1193 static tree
1194 alloc_max_size (void)
1196 if (!alloc_object_size_limit)
1198 alloc_object_size_limit = TYPE_MAX_VALUE (ssizetype);
1200 if (warn_alloc_size_limit)
1202 char *end = NULL;
1203 errno = 0;
1204 unsigned HOST_WIDE_INT unit = 1;
1205 unsigned HOST_WIDE_INT limit
1206 = strtoull (warn_alloc_size_limit, &end, 10);
1208 if (!errno)
1210 if (end && *end)
1212 /* Numeric option arguments are at most INT_MAX. Make it
1213 possible to specify a larger value by accepting common
1214 suffixes. */
1215 if (!strcmp (end, "kB"))
1216 unit = 1000;
1217 else if (!strcasecmp (end, "KiB") || strcmp (end, "KB"))
1218 unit = 1024;
1219 else if (!strcmp (end, "MB"))
1220 unit = 1000LU * 1000;
1221 else if (!strcasecmp (end, "MiB"))
1222 unit = 1024LU * 1024;
1223 else if (!strcasecmp (end, "GB"))
1224 unit = 1000LU * 1000 * 1000;
1225 else if (!strcasecmp (end, "GiB"))
1226 unit = 1024LU * 1024 * 1024;
1227 else if (!strcasecmp (end, "TB"))
1228 unit = 1000LU * 1000 * 1000 * 1000;
1229 else if (!strcasecmp (end, "TiB"))
1230 unit = 1024LU * 1024 * 1024 * 1024;
1231 else if (!strcasecmp (end, "PB"))
1232 unit = 1000LU * 1000 * 1000 * 1000 * 1000;
1233 else if (!strcasecmp (end, "PiB"))
1234 unit = 1024LU * 1024 * 1024 * 1024 * 1024;
1235 else if (!strcasecmp (end, "EB"))
1236 unit = 1000LU * 1000 * 1000 * 1000 * 1000 * 1000;
1237 else if (!strcasecmp (end, "EiB"))
1238 unit = 1024LU * 1024 * 1024 * 1024 * 1024 * 1024;
1239 else
1240 unit = 0;
1243 if (unit)
1244 alloc_object_size_limit
1245 = build_int_cst (ssizetype, limit * unit);
1249 return alloc_object_size_limit;
1252 /* Return true when EXP's range can be determined and set RANGE[] to it
1253 after adjusting it if necessary to make EXP a valid size argument to
1254 an allocation function declared with attribute alloc_size (whose
1255 argument may be signed), or to a string manipulation function like
1256 memset. */
1258 bool
1259 get_size_range (tree exp, tree range[2])
1261 if (tree_fits_uhwi_p (exp))
1263 /* EXP is a constant. */
1264 range[0] = range[1] = exp;
1265 return true;
1268 wide_int min, max;
1269 enum value_range_type range_type
1270 = (TREE_CODE (exp) == SSA_NAME
1271 ? get_range_info (exp, &min, &max) : VR_VARYING);
1273 if (range_type == VR_VARYING)
1275 /* No range information available. */
1276 range[0] = NULL_TREE;
1277 range[1] = NULL_TREE;
1278 return false;
1281 tree exptype = TREE_TYPE (exp);
1282 unsigned expprec = TYPE_PRECISION (exptype);
1283 wide_int wzero = wi::zero (expprec);
1284 wide_int wmaxval = wide_int (TYPE_MAX_VALUE (exptype));
1286 bool signed_p = !TYPE_UNSIGNED (exptype);
1288 if (range_type == VR_ANTI_RANGE)
1290 if (signed_p)
1292 if (wi::les_p (max, wzero))
1294 /* EXP is not in a strictly negative range. That means
1295 it must be in some (not necessarily strictly) positive
1296 range which includes zero. Since in signed to unsigned
1297 conversions negative values end up converted to large
1298 positive values, and otherwise they are not valid sizes,
1299 the resulting range is in both cases [0, TYPE_MAX]. */
1300 min = wzero;
1301 max = wmaxval;
1303 else if (wi::les_p (min - 1, wzero))
1305 /* EXP is not in a negative-positive range. That means EXP
1306 is either negative, or greater than max. Since negative
1307 sizes are invalid make the range [MAX + 1, TYPE_MAX]. */
1308 min = max + 1;
1309 max = wmaxval;
1311 else
1313 max = min - 1;
1314 min = wzero;
1317 else if (wi::eq_p (wzero, min - 1))
1319 /* EXP is unsigned and not in the range [1, MAX]. That means
1320 it's either zero or greater than MAX. Even though 0 would
1321 normally be detected by -Walloc-zero set the range to
1322 [MAX, TYPE_MAX] so that when MAX is greater than the limit
1323 the whole range is diagnosed. */
1324 min = max + 1;
1325 max = wmaxval;
1327 else
1329 max = min - 1;
1330 min = wzero;
1334 range[0] = wide_int_to_tree (exptype, min);
1335 range[1] = wide_int_to_tree (exptype, max);
1337 return true;
1340 /* Diagnose a call EXP to function FN decorated with attribute alloc_size
1341 whose argument numbers given by IDX with values given by ARGS exceed
1342 the maximum object size or cause an unsigned oveflow (wrapping) when
1343 multiplied. When ARGS[0] is null the function does nothing. ARGS[1]
1344 may be null for functions like malloc, and non-null for those like
1345 calloc that are decorated with a two-argument attribute alloc_size. */
1347 void
1348 maybe_warn_alloc_args_overflow (tree fn, tree exp, tree args[2], int idx[2])
1350 /* The range each of the (up to) two arguments is known to be in. */
1351 tree argrange[2][2] = { { NULL_TREE, NULL_TREE }, { NULL_TREE, NULL_TREE } };
1353 /* Maximum object size set by -Walloc-size-larger-than= or SIZE_MAX / 2. */
1354 tree maxobjsize = alloc_max_size ();
1356 location_t loc = EXPR_LOCATION (exp);
1358 bool warned = false;
1360 /* Validate each argument individually. */
1361 for (unsigned i = 0; i != 2 && args[i]; ++i)
1363 if (TREE_CODE (args[i]) == INTEGER_CST)
1365 argrange[i][0] = args[i];
1366 argrange[i][1] = args[i];
1368 if (tree_int_cst_lt (args[i], integer_zero_node))
1370 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1371 "%Kargument %i value %qE is negative",
1372 exp, idx[i] + 1, args[i]);
1374 else if (integer_zerop (args[i]))
1376 /* Avoid issuing -Walloc-zero for allocation functions other
1377 than __builtin_alloca that are declared with attribute
1378 returns_nonnull because there's no portability risk. This
1379 avoids warning for such calls to libiberty's xmalloc and
1380 friends.
1381 Also avoid issuing the warning for calls to function named
1382 "alloca". */
1383 if ((DECL_FUNCTION_CODE (fn) == BUILT_IN_ALLOCA
1384 && IDENTIFIER_LENGTH (DECL_NAME (fn)) != 6)
1385 || (DECL_FUNCTION_CODE (fn) != BUILT_IN_ALLOCA
1386 && !lookup_attribute ("returns_nonnull",
1387 TYPE_ATTRIBUTES (TREE_TYPE (fn)))))
1388 warned = warning_at (loc, OPT_Walloc_zero,
1389 "%Kargument %i value is zero",
1390 exp, idx[i] + 1);
1392 else if (tree_int_cst_lt (maxobjsize, args[i]))
1394 /* G++ emits calls to ::operator new[](SIZE_MAX) in C++98
1395 mode and with -fno-exceptions as a way to indicate array
1396 size overflow. There's no good way to detect C++98 here
1397 so avoid diagnosing these calls for all C++ modes. */
1398 if (i == 0
1399 && !args[1]
1400 && lang_GNU_CXX ()
1401 && DECL_IS_OPERATOR_NEW (fn)
1402 && integer_all_onesp (args[i]))
1403 continue;
1405 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1406 "%Kargument %i value %qE exceeds "
1407 "maximum object size %E",
1408 exp, idx[i] + 1, args[i], maxobjsize);
1411 else if (TREE_CODE (args[i]) == SSA_NAME
1412 && get_size_range (args[i], argrange[i]))
1414 /* Verify that the argument's range is not negative (including
1415 upper bound of zero). */
1416 if (tree_int_cst_lt (argrange[i][0], integer_zero_node)
1417 && tree_int_cst_le (argrange[i][1], integer_zero_node))
1419 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1420 "%Kargument %i range [%E, %E] is negative",
1421 exp, idx[i] + 1,
1422 argrange[i][0], argrange[i][1]);
1424 else if (tree_int_cst_lt (maxobjsize, argrange[i][0]))
1426 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1427 "%Kargument %i range [%E, %E] exceeds "
1428 "maximum object size %E",
1429 exp, idx[i] + 1,
1430 argrange[i][0], argrange[i][1],
1431 maxobjsize);
1436 if (!argrange[0])
1437 return;
1439 /* For a two-argument alloc_size, validate the product of the two
1440 arguments if both of their values or ranges are known. */
1441 if (!warned && tree_fits_uhwi_p (argrange[0][0])
1442 && argrange[1][0] && tree_fits_uhwi_p (argrange[1][0])
1443 && !integer_onep (argrange[0][0])
1444 && !integer_onep (argrange[1][0]))
1446 /* Check for overflow in the product of a function decorated with
1447 attribute alloc_size (X, Y). */
1448 unsigned szprec = TYPE_PRECISION (size_type_node);
1449 wide_int x = wi::to_wide (argrange[0][0], szprec);
1450 wide_int y = wi::to_wide (argrange[1][0], szprec);
1452 bool vflow;
1453 wide_int prod = wi::umul (x, y, &vflow);
1455 if (vflow)
1456 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1457 "%Kproduct %<%E * %E%> of arguments %i and %i "
1458 "exceeds %<SIZE_MAX%>",
1459 exp, argrange[0][0], argrange[1][0],
1460 idx[0] + 1, idx[1] + 1);
1461 else if (wi::ltu_p (wi::to_wide (maxobjsize, szprec), prod))
1462 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1463 "%Kproduct %<%E * %E%> of arguments %i and %i "
1464 "exceeds maximum object size %E",
1465 exp, argrange[0][0], argrange[1][0],
1466 idx[0] + 1, idx[1] + 1,
1467 maxobjsize);
1469 if (warned)
1471 /* Print the full range of each of the two arguments to make
1472 it clear when it is, in fact, in a range and not constant. */
1473 if (argrange[0][0] != argrange [0][1])
1474 inform (loc, "argument %i in the range [%E, %E]",
1475 idx[0] + 1, argrange[0][0], argrange[0][1]);
1476 if (argrange[1][0] != argrange [1][1])
1477 inform (loc, "argument %i in the range [%E, %E]",
1478 idx[1] + 1, argrange[1][0], argrange[1][1]);
1482 if (warned)
1484 location_t fnloc = DECL_SOURCE_LOCATION (fn);
1486 if (DECL_IS_BUILTIN (fn))
1487 inform (loc,
1488 "in a call to built-in allocation function %qD", fn);
1489 else
1490 inform (fnloc,
1491 "in a call to allocation function %qD declared here", fn);
1495 /* Issue an error if CALL_EXPR was flagged as requiring
1496 tall-call optimization. */
1498 static void
1499 maybe_complain_about_tail_call (tree call_expr, const char *reason)
1501 gcc_assert (TREE_CODE (call_expr) == CALL_EXPR);
1502 if (!CALL_EXPR_MUST_TAIL_CALL (call_expr))
1503 return;
1505 error_at (EXPR_LOCATION (call_expr), "cannot tail-call: %s", reason);
1508 /* Fill in ARGS_SIZE and ARGS array based on the parameters found in
1509 CALL_EXPR EXP.
1511 NUM_ACTUALS is the total number of parameters.
1513 N_NAMED_ARGS is the total number of named arguments.
1515 STRUCT_VALUE_ADDR_VALUE is the implicit argument for a struct return
1516 value, or null.
1518 FNDECL is the tree code for the target of this call (if known)
1520 ARGS_SO_FAR holds state needed by the target to know where to place
1521 the next argument.
1523 REG_PARM_STACK_SPACE is the number of bytes of stack space reserved
1524 for arguments which are passed in registers.
1526 OLD_STACK_LEVEL is a pointer to an rtx which olds the old stack level
1527 and may be modified by this routine.
1529 OLD_PENDING_ADJ, MUST_PREALLOCATE and FLAGS are pointers to integer
1530 flags which may be modified by this routine.
1532 MAY_TAILCALL is cleared if we encounter an invisible pass-by-reference
1533 that requires allocation of stack space.
1535 CALL_FROM_THUNK_P is true if this call is the jump from a thunk to
1536 the thunked-to function. */
1538 static void
1539 initialize_argument_information (int num_actuals ATTRIBUTE_UNUSED,
1540 struct arg_data *args,
1541 struct args_size *args_size,
1542 int n_named_args ATTRIBUTE_UNUSED,
1543 tree exp, tree struct_value_addr_value,
1544 tree fndecl, tree fntype,
1545 cumulative_args_t args_so_far,
1546 int reg_parm_stack_space,
1547 rtx *old_stack_level, int *old_pending_adj,
1548 int *must_preallocate, int *ecf_flags,
1549 bool *may_tailcall, bool call_from_thunk_p)
1551 CUMULATIVE_ARGS *args_so_far_pnt = get_cumulative_args (args_so_far);
1552 location_t loc = EXPR_LOCATION (exp);
1554 /* Count arg position in order args appear. */
1555 int argpos;
1557 int i;
1559 args_size->constant = 0;
1560 args_size->var = 0;
1562 bitmap_obstack_initialize (NULL);
1564 /* In this loop, we consider args in the order they are written.
1565 We fill up ARGS from the back. */
1567 i = num_actuals - 1;
1569 int j = i, ptr_arg = -1;
1570 call_expr_arg_iterator iter;
1571 tree arg;
1572 bitmap slots = NULL;
1574 if (struct_value_addr_value)
1576 args[j].tree_value = struct_value_addr_value;
1577 j--;
1579 /* If we pass structure address then we need to
1580 create bounds for it. Since created bounds is
1581 a call statement, we expand it right here to avoid
1582 fixing all other places where it may be expanded. */
1583 if (CALL_WITH_BOUNDS_P (exp))
1585 args[j].value = gen_reg_rtx (targetm.chkp_bound_mode ());
1586 args[j].tree_value
1587 = chkp_make_bounds_for_struct_addr (struct_value_addr_value);
1588 expand_expr_real (args[j].tree_value, args[j].value, VOIDmode,
1589 EXPAND_NORMAL, 0, false);
1590 args[j].pointer_arg = j + 1;
1591 j--;
1594 argpos = 0;
1595 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
1597 tree argtype = TREE_TYPE (arg);
1599 /* Remember last param with pointer and associate it
1600 with following pointer bounds. */
1601 if (CALL_WITH_BOUNDS_P (exp)
1602 && chkp_type_has_pointer (argtype))
1604 if (slots)
1605 BITMAP_FREE (slots);
1606 ptr_arg = j;
1607 if (!BOUNDED_TYPE_P (argtype))
1609 slots = BITMAP_ALLOC (NULL);
1610 chkp_find_bound_slots (argtype, slots);
1613 else if (CALL_WITH_BOUNDS_P (exp)
1614 && pass_by_reference (NULL, TYPE_MODE (argtype), argtype,
1615 argpos < n_named_args))
1617 if (slots)
1618 BITMAP_FREE (slots);
1619 ptr_arg = j;
1621 else if (POINTER_BOUNDS_TYPE_P (argtype))
1623 /* We expect bounds in instrumented calls only.
1624 Otherwise it is a sign we lost flag due to some optimization
1625 and may emit call args incorrectly. */
1626 gcc_assert (CALL_WITH_BOUNDS_P (exp));
1628 /* For structures look for the next available pointer. */
1629 if (ptr_arg != -1 && slots)
1631 unsigned bnd_no = bitmap_first_set_bit (slots);
1632 args[j].pointer_offset =
1633 bnd_no * POINTER_SIZE / BITS_PER_UNIT;
1635 bitmap_clear_bit (slots, bnd_no);
1637 /* Check we have no more pointers in the structure. */
1638 if (bitmap_empty_p (slots))
1639 BITMAP_FREE (slots);
1641 args[j].pointer_arg = ptr_arg;
1643 /* Check we covered all pointers in the previous
1644 non bounds arg. */
1645 if (!slots)
1646 ptr_arg = -1;
1648 else
1649 ptr_arg = -1;
1651 if (targetm.calls.split_complex_arg
1652 && argtype
1653 && TREE_CODE (argtype) == COMPLEX_TYPE
1654 && targetm.calls.split_complex_arg (argtype))
1656 tree subtype = TREE_TYPE (argtype);
1657 args[j].tree_value = build1 (REALPART_EXPR, subtype, arg);
1658 j--;
1659 args[j].tree_value = build1 (IMAGPART_EXPR, subtype, arg);
1661 else
1662 args[j].tree_value = arg;
1663 j--;
1664 argpos++;
1667 if (slots)
1668 BITMAP_FREE (slots);
1671 bitmap_obstack_release (NULL);
1673 /* Extract attribute alloc_size and if set, store the indices of
1674 the corresponding arguments in ALLOC_IDX, and then the actual
1675 argument(s) at those indices in ALLOC_ARGS. */
1676 int alloc_idx[2] = { -1, -1 };
1677 if (tree alloc_size
1678 = (fndecl ? lookup_attribute ("alloc_size",
1679 TYPE_ATTRIBUTES (TREE_TYPE (fndecl)))
1680 : NULL_TREE))
1682 tree args = TREE_VALUE (alloc_size);
1683 alloc_idx[0] = TREE_INT_CST_LOW (TREE_VALUE (args)) - 1;
1684 if (TREE_CHAIN (args))
1685 alloc_idx[1] = TREE_INT_CST_LOW (TREE_VALUE (TREE_CHAIN (args))) - 1;
1688 /* Array for up to the two attribute alloc_size arguments. */
1689 tree alloc_args[] = { NULL_TREE, NULL_TREE };
1691 /* I counts args in order (to be) pushed; ARGPOS counts in order written. */
1692 for (argpos = 0; argpos < num_actuals; i--, argpos++)
1694 tree type = TREE_TYPE (args[i].tree_value);
1695 int unsignedp;
1696 machine_mode mode;
1698 /* Replace erroneous argument with constant zero. */
1699 if (type == error_mark_node || !COMPLETE_TYPE_P (type))
1700 args[i].tree_value = integer_zero_node, type = integer_type_node;
1702 /* If TYPE is a transparent union or record, pass things the way
1703 we would pass the first field of the union or record. We have
1704 already verified that the modes are the same. */
1705 if ((TREE_CODE (type) == UNION_TYPE || TREE_CODE (type) == RECORD_TYPE)
1706 && TYPE_TRANSPARENT_AGGR (type))
1707 type = TREE_TYPE (first_field (type));
1709 /* Decide where to pass this arg.
1711 args[i].reg is nonzero if all or part is passed in registers.
1713 args[i].partial is nonzero if part but not all is passed in registers,
1714 and the exact value says how many bytes are passed in registers.
1716 args[i].pass_on_stack is nonzero if the argument must at least be
1717 computed on the stack. It may then be loaded back into registers
1718 if args[i].reg is nonzero.
1720 These decisions are driven by the FUNCTION_... macros and must agree
1721 with those made by function.c. */
1723 /* See if this argument should be passed by invisible reference. */
1724 if (pass_by_reference (args_so_far_pnt, TYPE_MODE (type),
1725 type, argpos < n_named_args))
1727 bool callee_copies;
1728 tree base = NULL_TREE;
1730 callee_copies
1731 = reference_callee_copied (args_so_far_pnt, TYPE_MODE (type),
1732 type, argpos < n_named_args);
1734 /* If we're compiling a thunk, pass through invisible references
1735 instead of making a copy. */
1736 if (call_from_thunk_p
1737 || (callee_copies
1738 && !TREE_ADDRESSABLE (type)
1739 && (base = get_base_address (args[i].tree_value))
1740 && TREE_CODE (base) != SSA_NAME
1741 && (!DECL_P (base) || MEM_P (DECL_RTL (base)))))
1743 /* We may have turned the parameter value into an SSA name.
1744 Go back to the original parameter so we can take the
1745 address. */
1746 if (TREE_CODE (args[i].tree_value) == SSA_NAME)
1748 gcc_assert (SSA_NAME_IS_DEFAULT_DEF (args[i].tree_value));
1749 args[i].tree_value = SSA_NAME_VAR (args[i].tree_value);
1750 gcc_assert (TREE_CODE (args[i].tree_value) == PARM_DECL);
1752 /* Argument setup code may have copied the value to register. We
1753 revert that optimization now because the tail call code must
1754 use the original location. */
1755 if (TREE_CODE (args[i].tree_value) == PARM_DECL
1756 && !MEM_P (DECL_RTL (args[i].tree_value))
1757 && DECL_INCOMING_RTL (args[i].tree_value)
1758 && MEM_P (DECL_INCOMING_RTL (args[i].tree_value)))
1759 set_decl_rtl (args[i].tree_value,
1760 DECL_INCOMING_RTL (args[i].tree_value));
1762 mark_addressable (args[i].tree_value);
1764 /* We can't use sibcalls if a callee-copied argument is
1765 stored in the current function's frame. */
1766 if (!call_from_thunk_p && DECL_P (base) && !TREE_STATIC (base))
1768 *may_tailcall = false;
1769 maybe_complain_about_tail_call (exp,
1770 "a callee-copied argument is"
1771 " stored in the current "
1772 " function's frame");
1775 args[i].tree_value = build_fold_addr_expr_loc (loc,
1776 args[i].tree_value);
1777 type = TREE_TYPE (args[i].tree_value);
1779 if (*ecf_flags & ECF_CONST)
1780 *ecf_flags &= ~(ECF_CONST | ECF_LOOPING_CONST_OR_PURE);
1782 else
1784 /* We make a copy of the object and pass the address to the
1785 function being called. */
1786 rtx copy;
1788 if (!COMPLETE_TYPE_P (type)
1789 || TREE_CODE (TYPE_SIZE_UNIT (type)) != INTEGER_CST
1790 || (flag_stack_check == GENERIC_STACK_CHECK
1791 && compare_tree_int (TYPE_SIZE_UNIT (type),
1792 STACK_CHECK_MAX_VAR_SIZE) > 0))
1794 /* This is a variable-sized object. Make space on the stack
1795 for it. */
1796 rtx size_rtx = expr_size (args[i].tree_value);
1798 if (*old_stack_level == 0)
1800 emit_stack_save (SAVE_BLOCK, old_stack_level);
1801 *old_pending_adj = pending_stack_adjust;
1802 pending_stack_adjust = 0;
1805 /* We can pass TRUE as the 4th argument because we just
1806 saved the stack pointer and will restore it right after
1807 the call. */
1808 copy = allocate_dynamic_stack_space (size_rtx,
1809 TYPE_ALIGN (type),
1810 TYPE_ALIGN (type),
1811 true);
1812 copy = gen_rtx_MEM (BLKmode, copy);
1813 set_mem_attributes (copy, type, 1);
1815 else
1816 copy = assign_temp (type, 1, 0);
1818 store_expr (args[i].tree_value, copy, 0, false, false);
1820 /* Just change the const function to pure and then let
1821 the next test clear the pure based on
1822 callee_copies. */
1823 if (*ecf_flags & ECF_CONST)
1825 *ecf_flags &= ~ECF_CONST;
1826 *ecf_flags |= ECF_PURE;
1829 if (!callee_copies && *ecf_flags & ECF_PURE)
1830 *ecf_flags &= ~(ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
1832 args[i].tree_value
1833 = build_fold_addr_expr_loc (loc, make_tree (type, copy));
1834 type = TREE_TYPE (args[i].tree_value);
1835 *may_tailcall = false;
1836 maybe_complain_about_tail_call (exp,
1837 "argument must be passed"
1838 " by copying");
1842 unsignedp = TYPE_UNSIGNED (type);
1843 mode = promote_function_mode (type, TYPE_MODE (type), &unsignedp,
1844 fndecl ? TREE_TYPE (fndecl) : fntype, 0);
1846 args[i].unsignedp = unsignedp;
1847 args[i].mode = mode;
1849 args[i].reg = targetm.calls.function_arg (args_so_far, mode, type,
1850 argpos < n_named_args);
1852 if (args[i].reg && CONST_INT_P (args[i].reg))
1854 args[i].special_slot = args[i].reg;
1855 args[i].reg = NULL;
1858 /* If this is a sibling call and the machine has register windows, the
1859 register window has to be unwinded before calling the routine, so
1860 arguments have to go into the incoming registers. */
1861 if (targetm.calls.function_incoming_arg != targetm.calls.function_arg)
1862 args[i].tail_call_reg
1863 = targetm.calls.function_incoming_arg (args_so_far, mode, type,
1864 argpos < n_named_args);
1865 else
1866 args[i].tail_call_reg = args[i].reg;
1868 if (args[i].reg)
1869 args[i].partial
1870 = targetm.calls.arg_partial_bytes (args_so_far, mode, type,
1871 argpos < n_named_args);
1873 args[i].pass_on_stack = targetm.calls.must_pass_in_stack (mode, type);
1875 /* If FUNCTION_ARG returned a (parallel [(expr_list (nil) ...) ...]),
1876 it means that we are to pass this arg in the register(s) designated
1877 by the PARALLEL, but also to pass it in the stack. */
1878 if (args[i].reg && GET_CODE (args[i].reg) == PARALLEL
1879 && XEXP (XVECEXP (args[i].reg, 0, 0), 0) == 0)
1880 args[i].pass_on_stack = 1;
1882 /* If this is an addressable type, we must preallocate the stack
1883 since we must evaluate the object into its final location.
1885 If this is to be passed in both registers and the stack, it is simpler
1886 to preallocate. */
1887 if (TREE_ADDRESSABLE (type)
1888 || (args[i].pass_on_stack && args[i].reg != 0))
1889 *must_preallocate = 1;
1891 /* No stack allocation and padding for bounds. */
1892 if (POINTER_BOUNDS_P (args[i].tree_value))
1894 /* Compute the stack-size of this argument. */
1895 else if (args[i].reg == 0 || args[i].partial != 0
1896 || reg_parm_stack_space > 0
1897 || args[i].pass_on_stack)
1898 locate_and_pad_parm (mode, type,
1899 #ifdef STACK_PARMS_IN_REG_PARM_AREA
1901 #else
1902 args[i].reg != 0,
1903 #endif
1904 reg_parm_stack_space,
1905 args[i].pass_on_stack ? 0 : args[i].partial,
1906 fndecl, args_size, &args[i].locate);
1907 #ifdef BLOCK_REG_PADDING
1908 else
1909 /* The argument is passed entirely in registers. See at which
1910 end it should be padded. */
1911 args[i].locate.where_pad =
1912 BLOCK_REG_PADDING (mode, type,
1913 int_size_in_bytes (type) <= UNITS_PER_WORD);
1914 #endif
1916 /* Update ARGS_SIZE, the total stack space for args so far. */
1918 args_size->constant += args[i].locate.size.constant;
1919 if (args[i].locate.size.var)
1920 ADD_PARM_SIZE (*args_size, args[i].locate.size.var);
1922 /* Increment ARGS_SO_FAR, which has info about which arg-registers
1923 have been used, etc. */
1925 targetm.calls.function_arg_advance (args_so_far, TYPE_MODE (type),
1926 type, argpos < n_named_args);
1928 /* Store argument values for functions decorated with attribute
1929 alloc_size. */
1930 if (argpos == alloc_idx[0])
1931 alloc_args[0] = args[i].tree_value;
1932 else if (argpos == alloc_idx[1])
1933 alloc_args[1] = args[i].tree_value;
1936 if (alloc_args[0])
1938 /* Check the arguments of functions decorated with attribute
1939 alloc_size. */
1940 maybe_warn_alloc_args_overflow (fndecl, exp, alloc_args, alloc_idx);
1944 /* Update ARGS_SIZE to contain the total size for the argument block.
1945 Return the original constant component of the argument block's size.
1947 REG_PARM_STACK_SPACE holds the number of bytes of stack space reserved
1948 for arguments passed in registers. */
1950 static int
1951 compute_argument_block_size (int reg_parm_stack_space,
1952 struct args_size *args_size,
1953 tree fndecl ATTRIBUTE_UNUSED,
1954 tree fntype ATTRIBUTE_UNUSED,
1955 int preferred_stack_boundary ATTRIBUTE_UNUSED)
1957 int unadjusted_args_size = args_size->constant;
1959 /* For accumulate outgoing args mode we don't need to align, since the frame
1960 will be already aligned. Align to STACK_BOUNDARY in order to prevent
1961 backends from generating misaligned frame sizes. */
1962 if (ACCUMULATE_OUTGOING_ARGS && preferred_stack_boundary > STACK_BOUNDARY)
1963 preferred_stack_boundary = STACK_BOUNDARY;
1965 /* Compute the actual size of the argument block required. The variable
1966 and constant sizes must be combined, the size may have to be rounded,
1967 and there may be a minimum required size. */
1969 if (args_size->var)
1971 args_size->var = ARGS_SIZE_TREE (*args_size);
1972 args_size->constant = 0;
1974 preferred_stack_boundary /= BITS_PER_UNIT;
1975 if (preferred_stack_boundary > 1)
1977 /* We don't handle this case yet. To handle it correctly we have
1978 to add the delta, round and subtract the delta.
1979 Currently no machine description requires this support. */
1980 gcc_assert (!(stack_pointer_delta & (preferred_stack_boundary - 1)));
1981 args_size->var = round_up (args_size->var, preferred_stack_boundary);
1984 if (reg_parm_stack_space > 0)
1986 args_size->var
1987 = size_binop (MAX_EXPR, args_size->var,
1988 ssize_int (reg_parm_stack_space));
1990 /* The area corresponding to register parameters is not to count in
1991 the size of the block we need. So make the adjustment. */
1992 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
1993 args_size->var
1994 = size_binop (MINUS_EXPR, args_size->var,
1995 ssize_int (reg_parm_stack_space));
1998 else
2000 preferred_stack_boundary /= BITS_PER_UNIT;
2001 if (preferred_stack_boundary < 1)
2002 preferred_stack_boundary = 1;
2003 args_size->constant = (((args_size->constant
2004 + stack_pointer_delta
2005 + preferred_stack_boundary - 1)
2006 / preferred_stack_boundary
2007 * preferred_stack_boundary)
2008 - stack_pointer_delta);
2010 args_size->constant = MAX (args_size->constant,
2011 reg_parm_stack_space);
2013 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
2014 args_size->constant -= reg_parm_stack_space;
2016 return unadjusted_args_size;
2019 /* Precompute parameters as needed for a function call.
2021 FLAGS is mask of ECF_* constants.
2023 NUM_ACTUALS is the number of arguments.
2025 ARGS is an array containing information for each argument; this
2026 routine fills in the INITIAL_VALUE and VALUE fields for each
2027 precomputed argument. */
2029 static void
2030 precompute_arguments (int num_actuals, struct arg_data *args)
2032 int i;
2034 /* If this is a libcall, then precompute all arguments so that we do not
2035 get extraneous instructions emitted as part of the libcall sequence. */
2037 /* If we preallocated the stack space, and some arguments must be passed
2038 on the stack, then we must precompute any parameter which contains a
2039 function call which will store arguments on the stack.
2040 Otherwise, evaluating the parameter may clobber previous parameters
2041 which have already been stored into the stack. (we have code to avoid
2042 such case by saving the outgoing stack arguments, but it results in
2043 worse code) */
2044 if (!ACCUMULATE_OUTGOING_ARGS)
2045 return;
2047 for (i = 0; i < num_actuals; i++)
2049 tree type;
2050 machine_mode mode;
2052 if (TREE_CODE (args[i].tree_value) != CALL_EXPR)
2053 continue;
2055 /* If this is an addressable type, we cannot pre-evaluate it. */
2056 type = TREE_TYPE (args[i].tree_value);
2057 gcc_assert (!TREE_ADDRESSABLE (type));
2059 args[i].initial_value = args[i].value
2060 = expand_normal (args[i].tree_value);
2062 mode = TYPE_MODE (type);
2063 if (mode != args[i].mode)
2065 int unsignedp = args[i].unsignedp;
2066 args[i].value
2067 = convert_modes (args[i].mode, mode,
2068 args[i].value, args[i].unsignedp);
2070 /* CSE will replace this only if it contains args[i].value
2071 pseudo, so convert it down to the declared mode using
2072 a SUBREG. */
2073 if (REG_P (args[i].value)
2074 && GET_MODE_CLASS (args[i].mode) == MODE_INT
2075 && promote_mode (type, mode, &unsignedp) != args[i].mode)
2077 args[i].initial_value
2078 = gen_lowpart_SUBREG (mode, args[i].value);
2079 SUBREG_PROMOTED_VAR_P (args[i].initial_value) = 1;
2080 SUBREG_PROMOTED_SET (args[i].initial_value, args[i].unsignedp);
2086 /* Given the current state of MUST_PREALLOCATE and information about
2087 arguments to a function call in NUM_ACTUALS, ARGS and ARGS_SIZE,
2088 compute and return the final value for MUST_PREALLOCATE. */
2090 static int
2091 finalize_must_preallocate (int must_preallocate, int num_actuals,
2092 struct arg_data *args, struct args_size *args_size)
2094 /* See if we have or want to preallocate stack space.
2096 If we would have to push a partially-in-regs parm
2097 before other stack parms, preallocate stack space instead.
2099 If the size of some parm is not a multiple of the required stack
2100 alignment, we must preallocate.
2102 If the total size of arguments that would otherwise create a copy in
2103 a temporary (such as a CALL) is more than half the total argument list
2104 size, preallocation is faster.
2106 Another reason to preallocate is if we have a machine (like the m88k)
2107 where stack alignment is required to be maintained between every
2108 pair of insns, not just when the call is made. However, we assume here
2109 that such machines either do not have push insns (and hence preallocation
2110 would occur anyway) or the problem is taken care of with
2111 PUSH_ROUNDING. */
2113 if (! must_preallocate)
2115 int partial_seen = 0;
2116 int copy_to_evaluate_size = 0;
2117 int i;
2119 for (i = 0; i < num_actuals && ! must_preallocate; i++)
2121 if (args[i].partial > 0 && ! args[i].pass_on_stack)
2122 partial_seen = 1;
2123 else if (partial_seen && args[i].reg == 0)
2124 must_preallocate = 1;
2125 /* We preallocate in case there are bounds passed
2126 in the bounds table to have precomputed address
2127 for bounds association. */
2128 else if (POINTER_BOUNDS_P (args[i].tree_value)
2129 && !args[i].reg)
2130 must_preallocate = 1;
2132 if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode
2133 && (TREE_CODE (args[i].tree_value) == CALL_EXPR
2134 || TREE_CODE (args[i].tree_value) == TARGET_EXPR
2135 || TREE_CODE (args[i].tree_value) == COND_EXPR
2136 || TREE_ADDRESSABLE (TREE_TYPE (args[i].tree_value))))
2137 copy_to_evaluate_size
2138 += int_size_in_bytes (TREE_TYPE (args[i].tree_value));
2141 if (copy_to_evaluate_size * 2 >= args_size->constant
2142 && args_size->constant > 0)
2143 must_preallocate = 1;
2145 return must_preallocate;
2148 /* If we preallocated stack space, compute the address of each argument
2149 and store it into the ARGS array.
2151 We need not ensure it is a valid memory address here; it will be
2152 validized when it is used.
2154 ARGBLOCK is an rtx for the address of the outgoing arguments. */
2156 static void
2157 compute_argument_addresses (struct arg_data *args, rtx argblock, int num_actuals)
2159 if (argblock)
2161 rtx arg_reg = argblock;
2162 int i, arg_offset = 0;
2164 if (GET_CODE (argblock) == PLUS)
2165 arg_reg = XEXP (argblock, 0), arg_offset = INTVAL (XEXP (argblock, 1));
2167 for (i = 0; i < num_actuals; i++)
2169 rtx offset = ARGS_SIZE_RTX (args[i].locate.offset);
2170 rtx slot_offset = ARGS_SIZE_RTX (args[i].locate.slot_offset);
2171 rtx addr;
2172 unsigned int align, boundary;
2173 unsigned int units_on_stack = 0;
2174 machine_mode partial_mode = VOIDmode;
2176 /* Skip this parm if it will not be passed on the stack. */
2177 if (! args[i].pass_on_stack
2178 && args[i].reg != 0
2179 && args[i].partial == 0)
2180 continue;
2182 /* Pointer Bounds are never passed on the stack. */
2183 if (POINTER_BOUNDS_P (args[i].tree_value))
2184 continue;
2186 if (CONST_INT_P (offset))
2187 addr = plus_constant (Pmode, arg_reg, INTVAL (offset));
2188 else
2189 addr = gen_rtx_PLUS (Pmode, arg_reg, offset);
2191 addr = plus_constant (Pmode, addr, arg_offset);
2193 if (args[i].partial != 0)
2195 /* Only part of the parameter is being passed on the stack.
2196 Generate a simple memory reference of the correct size. */
2197 units_on_stack = args[i].locate.size.constant;
2198 partial_mode = mode_for_size (units_on_stack * BITS_PER_UNIT,
2199 MODE_INT, 1);
2200 args[i].stack = gen_rtx_MEM (partial_mode, addr);
2201 set_mem_size (args[i].stack, units_on_stack);
2203 else
2205 args[i].stack = gen_rtx_MEM (args[i].mode, addr);
2206 set_mem_attributes (args[i].stack,
2207 TREE_TYPE (args[i].tree_value), 1);
2209 align = BITS_PER_UNIT;
2210 boundary = args[i].locate.boundary;
2211 if (args[i].locate.where_pad != downward)
2212 align = boundary;
2213 else if (CONST_INT_P (offset))
2215 align = INTVAL (offset) * BITS_PER_UNIT | boundary;
2216 align = least_bit_hwi (align);
2218 set_mem_align (args[i].stack, align);
2220 if (CONST_INT_P (slot_offset))
2221 addr = plus_constant (Pmode, arg_reg, INTVAL (slot_offset));
2222 else
2223 addr = gen_rtx_PLUS (Pmode, arg_reg, slot_offset);
2225 addr = plus_constant (Pmode, addr, arg_offset);
2227 if (args[i].partial != 0)
2229 /* Only part of the parameter is being passed on the stack.
2230 Generate a simple memory reference of the correct size.
2232 args[i].stack_slot = gen_rtx_MEM (partial_mode, addr);
2233 set_mem_size (args[i].stack_slot, units_on_stack);
2235 else
2237 args[i].stack_slot = gen_rtx_MEM (args[i].mode, addr);
2238 set_mem_attributes (args[i].stack_slot,
2239 TREE_TYPE (args[i].tree_value), 1);
2241 set_mem_align (args[i].stack_slot, args[i].locate.boundary);
2243 /* Function incoming arguments may overlap with sibling call
2244 outgoing arguments and we cannot allow reordering of reads
2245 from function arguments with stores to outgoing arguments
2246 of sibling calls. */
2247 set_mem_alias_set (args[i].stack, 0);
2248 set_mem_alias_set (args[i].stack_slot, 0);
2253 /* Given a FNDECL and EXP, return an rtx suitable for use as a target address
2254 in a call instruction.
2256 FNDECL is the tree node for the target function. For an indirect call
2257 FNDECL will be NULL_TREE.
2259 ADDR is the operand 0 of CALL_EXPR for this call. */
2261 static rtx
2262 rtx_for_function_call (tree fndecl, tree addr)
2264 rtx funexp;
2266 /* Get the function to call, in the form of RTL. */
2267 if (fndecl)
2269 if (!TREE_USED (fndecl) && fndecl != current_function_decl)
2270 TREE_USED (fndecl) = 1;
2272 /* Get a SYMBOL_REF rtx for the function address. */
2273 funexp = XEXP (DECL_RTL (fndecl), 0);
2275 else
2276 /* Generate an rtx (probably a pseudo-register) for the address. */
2278 push_temp_slots ();
2279 funexp = expand_normal (addr);
2280 pop_temp_slots (); /* FUNEXP can't be BLKmode. */
2282 return funexp;
2285 /* Internal state for internal_arg_pointer_based_exp and its helpers. */
2286 static struct
2288 /* Last insn that has been scanned by internal_arg_pointer_based_exp_scan,
2289 or NULL_RTX if none has been scanned yet. */
2290 rtx_insn *scan_start;
2291 /* Vector indexed by REGNO - FIRST_PSEUDO_REGISTER, recording if a pseudo is
2292 based on crtl->args.internal_arg_pointer. The element is NULL_RTX if the
2293 pseudo isn't based on it, a CONST_INT offset if the pseudo is based on it
2294 with fixed offset, or PC if this is with variable or unknown offset. */
2295 vec<rtx> cache;
2296 } internal_arg_pointer_exp_state;
2298 static rtx internal_arg_pointer_based_exp (const_rtx, bool);
2300 /* Helper function for internal_arg_pointer_based_exp. Scan insns in
2301 the tail call sequence, starting with first insn that hasn't been
2302 scanned yet, and note for each pseudo on the LHS whether it is based
2303 on crtl->args.internal_arg_pointer or not, and what offset from that
2304 that pointer it has. */
2306 static void
2307 internal_arg_pointer_based_exp_scan (void)
2309 rtx_insn *insn, *scan_start = internal_arg_pointer_exp_state.scan_start;
2311 if (scan_start == NULL_RTX)
2312 insn = get_insns ();
2313 else
2314 insn = NEXT_INSN (scan_start);
2316 while (insn)
2318 rtx set = single_set (insn);
2319 if (set && REG_P (SET_DEST (set)) && !HARD_REGISTER_P (SET_DEST (set)))
2321 rtx val = NULL_RTX;
2322 unsigned int idx = REGNO (SET_DEST (set)) - FIRST_PSEUDO_REGISTER;
2323 /* Punt on pseudos set multiple times. */
2324 if (idx < internal_arg_pointer_exp_state.cache.length ()
2325 && (internal_arg_pointer_exp_state.cache[idx]
2326 != NULL_RTX))
2327 val = pc_rtx;
2328 else
2329 val = internal_arg_pointer_based_exp (SET_SRC (set), false);
2330 if (val != NULL_RTX)
2332 if (idx >= internal_arg_pointer_exp_state.cache.length ())
2333 internal_arg_pointer_exp_state.cache
2334 .safe_grow_cleared (idx + 1);
2335 internal_arg_pointer_exp_state.cache[idx] = val;
2338 if (NEXT_INSN (insn) == NULL_RTX)
2339 scan_start = insn;
2340 insn = NEXT_INSN (insn);
2343 internal_arg_pointer_exp_state.scan_start = scan_start;
2346 /* Compute whether RTL is based on crtl->args.internal_arg_pointer. Return
2347 NULL_RTX if RTL isn't based on it, a CONST_INT offset if RTL is based on
2348 it with fixed offset, or PC if this is with variable or unknown offset.
2349 TOPLEVEL is true if the function is invoked at the topmost level. */
2351 static rtx
2352 internal_arg_pointer_based_exp (const_rtx rtl, bool toplevel)
2354 if (CONSTANT_P (rtl))
2355 return NULL_RTX;
2357 if (rtl == crtl->args.internal_arg_pointer)
2358 return const0_rtx;
2360 if (REG_P (rtl) && HARD_REGISTER_P (rtl))
2361 return NULL_RTX;
2363 if (GET_CODE (rtl) == PLUS && CONST_INT_P (XEXP (rtl, 1)))
2365 rtx val = internal_arg_pointer_based_exp (XEXP (rtl, 0), toplevel);
2366 if (val == NULL_RTX || val == pc_rtx)
2367 return val;
2368 return plus_constant (Pmode, val, INTVAL (XEXP (rtl, 1)));
2371 /* When called at the topmost level, scan pseudo assignments in between the
2372 last scanned instruction in the tail call sequence and the latest insn
2373 in that sequence. */
2374 if (toplevel)
2375 internal_arg_pointer_based_exp_scan ();
2377 if (REG_P (rtl))
2379 unsigned int idx = REGNO (rtl) - FIRST_PSEUDO_REGISTER;
2380 if (idx < internal_arg_pointer_exp_state.cache.length ())
2381 return internal_arg_pointer_exp_state.cache[idx];
2383 return NULL_RTX;
2386 subrtx_iterator::array_type array;
2387 FOR_EACH_SUBRTX (iter, array, rtl, NONCONST)
2389 const_rtx x = *iter;
2390 if (REG_P (x) && internal_arg_pointer_based_exp (x, false) != NULL_RTX)
2391 return pc_rtx;
2392 if (MEM_P (x))
2393 iter.skip_subrtxes ();
2396 return NULL_RTX;
2399 /* Return true if and only if SIZE storage units (usually bytes)
2400 starting from address ADDR overlap with already clobbered argument
2401 area. This function is used to determine if we should give up a
2402 sibcall. */
2404 static bool
2405 mem_overlaps_already_clobbered_arg_p (rtx addr, unsigned HOST_WIDE_INT size)
2407 HOST_WIDE_INT i;
2408 rtx val;
2410 if (bitmap_empty_p (stored_args_map))
2411 return false;
2412 val = internal_arg_pointer_based_exp (addr, true);
2413 if (val == NULL_RTX)
2414 return false;
2415 else if (val == pc_rtx)
2416 return true;
2417 else
2418 i = INTVAL (val);
2420 if (STACK_GROWS_DOWNWARD)
2421 i -= crtl->args.pretend_args_size;
2422 else
2423 i += crtl->args.pretend_args_size;
2426 if (ARGS_GROW_DOWNWARD)
2427 i = -i - size;
2429 if (size > 0)
2431 unsigned HOST_WIDE_INT k;
2433 for (k = 0; k < size; k++)
2434 if (i + k < SBITMAP_SIZE (stored_args_map)
2435 && bitmap_bit_p (stored_args_map, i + k))
2436 return true;
2439 return false;
2442 /* Do the register loads required for any wholly-register parms or any
2443 parms which are passed both on the stack and in a register. Their
2444 expressions were already evaluated.
2446 Mark all register-parms as living through the call, putting these USE
2447 insns in the CALL_INSN_FUNCTION_USAGE field.
2449 When IS_SIBCALL, perform the check_sibcall_argument_overlap
2450 checking, setting *SIBCALL_FAILURE if appropriate. */
2452 static void
2453 load_register_parameters (struct arg_data *args, int num_actuals,
2454 rtx *call_fusage, int flags, int is_sibcall,
2455 int *sibcall_failure)
2457 int i, j;
2459 for (i = 0; i < num_actuals; i++)
2461 rtx reg = ((flags & ECF_SIBCALL)
2462 ? args[i].tail_call_reg : args[i].reg);
2463 if (reg)
2465 int partial = args[i].partial;
2466 int nregs;
2467 int size = 0;
2468 rtx_insn *before_arg = get_last_insn ();
2469 /* Set non-negative if we must move a word at a time, even if
2470 just one word (e.g, partial == 4 && mode == DFmode). Set
2471 to -1 if we just use a normal move insn. This value can be
2472 zero if the argument is a zero size structure. */
2473 nregs = -1;
2474 if (GET_CODE (reg) == PARALLEL)
2476 else if (partial)
2478 gcc_assert (partial % UNITS_PER_WORD == 0);
2479 nregs = partial / UNITS_PER_WORD;
2481 else if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode)
2483 size = int_size_in_bytes (TREE_TYPE (args[i].tree_value));
2484 nregs = (size + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD;
2486 else
2487 size = GET_MODE_SIZE (args[i].mode);
2489 /* Handle calls that pass values in multiple non-contiguous
2490 locations. The Irix 6 ABI has examples of this. */
2492 if (GET_CODE (reg) == PARALLEL)
2493 emit_group_move (reg, args[i].parallel_value);
2495 /* If simple case, just do move. If normal partial, store_one_arg
2496 has already loaded the register for us. In all other cases,
2497 load the register(s) from memory. */
2499 else if (nregs == -1)
2501 emit_move_insn (reg, args[i].value);
2502 #ifdef BLOCK_REG_PADDING
2503 /* Handle case where we have a value that needs shifting
2504 up to the msb. eg. a QImode value and we're padding
2505 upward on a BYTES_BIG_ENDIAN machine. */
2506 if (size < UNITS_PER_WORD
2507 && (args[i].locate.where_pad
2508 == (BYTES_BIG_ENDIAN ? upward : downward)))
2510 rtx x;
2511 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
2513 /* Assigning REG here rather than a temp makes CALL_FUSAGE
2514 report the whole reg as used. Strictly speaking, the
2515 call only uses SIZE bytes at the msb end, but it doesn't
2516 seem worth generating rtl to say that. */
2517 reg = gen_rtx_REG (word_mode, REGNO (reg));
2518 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1);
2519 if (x != reg)
2520 emit_move_insn (reg, x);
2522 #endif
2525 /* If we have pre-computed the values to put in the registers in
2526 the case of non-aligned structures, copy them in now. */
2528 else if (args[i].n_aligned_regs != 0)
2529 for (j = 0; j < args[i].n_aligned_regs; j++)
2530 emit_move_insn (gen_rtx_REG (word_mode, REGNO (reg) + j),
2531 args[i].aligned_regs[j]);
2533 else if (partial == 0 || args[i].pass_on_stack)
2535 rtx mem = validize_mem (copy_rtx (args[i].value));
2537 /* Check for overlap with already clobbered argument area,
2538 providing that this has non-zero size. */
2539 if (is_sibcall
2540 && size != 0
2541 && (mem_overlaps_already_clobbered_arg_p
2542 (XEXP (args[i].value, 0), size)))
2543 *sibcall_failure = 1;
2545 if (size % UNITS_PER_WORD == 0
2546 || MEM_ALIGN (mem) % BITS_PER_WORD == 0)
2547 move_block_to_reg (REGNO (reg), mem, nregs, args[i].mode);
2548 else
2550 if (nregs > 1)
2551 move_block_to_reg (REGNO (reg), mem, nregs - 1,
2552 args[i].mode);
2553 rtx dest = gen_rtx_REG (word_mode, REGNO (reg) + nregs - 1);
2554 unsigned int bitoff = (nregs - 1) * BITS_PER_WORD;
2555 unsigned int bitsize = size * BITS_PER_UNIT - bitoff;
2556 rtx x = extract_bit_field (mem, bitsize, bitoff, 1, dest,
2557 word_mode, word_mode, false);
2558 if (BYTES_BIG_ENDIAN)
2559 x = expand_shift (LSHIFT_EXPR, word_mode, x,
2560 BITS_PER_WORD - bitsize, dest, 1);
2561 if (x != dest)
2562 emit_move_insn (dest, x);
2565 /* Handle a BLKmode that needs shifting. */
2566 if (nregs == 1 && size < UNITS_PER_WORD
2567 #ifdef BLOCK_REG_PADDING
2568 && args[i].locate.where_pad == downward
2569 #else
2570 && BYTES_BIG_ENDIAN
2571 #endif
2574 rtx dest = gen_rtx_REG (word_mode, REGNO (reg));
2575 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
2576 enum tree_code dir = (BYTES_BIG_ENDIAN
2577 ? RSHIFT_EXPR : LSHIFT_EXPR);
2578 rtx x;
2580 x = expand_shift (dir, word_mode, dest, shift, dest, 1);
2581 if (x != dest)
2582 emit_move_insn (dest, x);
2586 /* When a parameter is a block, and perhaps in other cases, it is
2587 possible that it did a load from an argument slot that was
2588 already clobbered. */
2589 if (is_sibcall
2590 && check_sibcall_argument_overlap (before_arg, &args[i], 0))
2591 *sibcall_failure = 1;
2593 /* Handle calls that pass values in multiple non-contiguous
2594 locations. The Irix 6 ABI has examples of this. */
2595 if (GET_CODE (reg) == PARALLEL)
2596 use_group_regs (call_fusage, reg);
2597 else if (nregs == -1)
2598 use_reg_mode (call_fusage, reg,
2599 TYPE_MODE (TREE_TYPE (args[i].tree_value)));
2600 else if (nregs > 0)
2601 use_regs (call_fusage, REGNO (reg), nregs);
2606 /* We need to pop PENDING_STACK_ADJUST bytes. But, if the arguments
2607 wouldn't fill up an even multiple of PREFERRED_UNIT_STACK_BOUNDARY
2608 bytes, then we would need to push some additional bytes to pad the
2609 arguments. So, we compute an adjust to the stack pointer for an
2610 amount that will leave the stack under-aligned by UNADJUSTED_ARGS_SIZE
2611 bytes. Then, when the arguments are pushed the stack will be perfectly
2612 aligned. ARGS_SIZE->CONSTANT is set to the number of bytes that should
2613 be popped after the call. Returns the adjustment. */
2615 static int
2616 combine_pending_stack_adjustment_and_call (int unadjusted_args_size,
2617 struct args_size *args_size,
2618 unsigned int preferred_unit_stack_boundary)
2620 /* The number of bytes to pop so that the stack will be
2621 under-aligned by UNADJUSTED_ARGS_SIZE bytes. */
2622 HOST_WIDE_INT adjustment;
2623 /* The alignment of the stack after the arguments are pushed, if we
2624 just pushed the arguments without adjust the stack here. */
2625 unsigned HOST_WIDE_INT unadjusted_alignment;
2627 unadjusted_alignment
2628 = ((stack_pointer_delta + unadjusted_args_size)
2629 % preferred_unit_stack_boundary);
2631 /* We want to get rid of as many of the PENDING_STACK_ADJUST bytes
2632 as possible -- leaving just enough left to cancel out the
2633 UNADJUSTED_ALIGNMENT. In other words, we want to ensure that the
2634 PENDING_STACK_ADJUST is non-negative, and congruent to
2635 -UNADJUSTED_ALIGNMENT modulo the PREFERRED_UNIT_STACK_BOUNDARY. */
2637 /* Begin by trying to pop all the bytes. */
2638 unadjusted_alignment
2639 = (unadjusted_alignment
2640 - (pending_stack_adjust % preferred_unit_stack_boundary));
2641 adjustment = pending_stack_adjust;
2642 /* Push enough additional bytes that the stack will be aligned
2643 after the arguments are pushed. */
2644 if (preferred_unit_stack_boundary > 1)
2646 if (unadjusted_alignment > 0)
2647 adjustment -= preferred_unit_stack_boundary - unadjusted_alignment;
2648 else
2649 adjustment += unadjusted_alignment;
2652 /* Now, sets ARGS_SIZE->CONSTANT so that we pop the right number of
2653 bytes after the call. The right number is the entire
2654 PENDING_STACK_ADJUST less our ADJUSTMENT plus the amount required
2655 by the arguments in the first place. */
2656 args_size->constant
2657 = pending_stack_adjust - adjustment + unadjusted_args_size;
2659 return adjustment;
2662 /* Scan X expression if it does not dereference any argument slots
2663 we already clobbered by tail call arguments (as noted in stored_args_map
2664 bitmap).
2665 Return nonzero if X expression dereferences such argument slots,
2666 zero otherwise. */
2668 static int
2669 check_sibcall_argument_overlap_1 (rtx x)
2671 RTX_CODE code;
2672 int i, j;
2673 const char *fmt;
2675 if (x == NULL_RTX)
2676 return 0;
2678 code = GET_CODE (x);
2680 /* We need not check the operands of the CALL expression itself. */
2681 if (code == CALL)
2682 return 0;
2684 if (code == MEM)
2685 return mem_overlaps_already_clobbered_arg_p (XEXP (x, 0),
2686 GET_MODE_SIZE (GET_MODE (x)));
2688 /* Scan all subexpressions. */
2689 fmt = GET_RTX_FORMAT (code);
2690 for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++)
2692 if (*fmt == 'e')
2694 if (check_sibcall_argument_overlap_1 (XEXP (x, i)))
2695 return 1;
2697 else if (*fmt == 'E')
2699 for (j = 0; j < XVECLEN (x, i); j++)
2700 if (check_sibcall_argument_overlap_1 (XVECEXP (x, i, j)))
2701 return 1;
2704 return 0;
2707 /* Scan sequence after INSN if it does not dereference any argument slots
2708 we already clobbered by tail call arguments (as noted in stored_args_map
2709 bitmap). If MARK_STORED_ARGS_MAP, add stack slots for ARG to
2710 stored_args_map bitmap afterwards (when ARG is a register MARK_STORED_ARGS_MAP
2711 should be 0). Return nonzero if sequence after INSN dereferences such argument
2712 slots, zero otherwise. */
2714 static int
2715 check_sibcall_argument_overlap (rtx_insn *insn, struct arg_data *arg,
2716 int mark_stored_args_map)
2718 int low, high;
2720 if (insn == NULL_RTX)
2721 insn = get_insns ();
2722 else
2723 insn = NEXT_INSN (insn);
2725 for (; insn; insn = NEXT_INSN (insn))
2726 if (INSN_P (insn)
2727 && check_sibcall_argument_overlap_1 (PATTERN (insn)))
2728 break;
2730 if (mark_stored_args_map)
2732 if (ARGS_GROW_DOWNWARD)
2733 low = -arg->locate.slot_offset.constant - arg->locate.size.constant;
2734 else
2735 low = arg->locate.slot_offset.constant;
2737 for (high = low + arg->locate.size.constant; low < high; low++)
2738 bitmap_set_bit (stored_args_map, low);
2740 return insn != NULL_RTX;
2743 /* Given that a function returns a value of mode MODE at the most
2744 significant end of hard register VALUE, shift VALUE left or right
2745 as specified by LEFT_P. Return true if some action was needed. */
2747 bool
2748 shift_return_value (machine_mode mode, bool left_p, rtx value)
2750 HOST_WIDE_INT shift;
2752 gcc_assert (REG_P (value) && HARD_REGISTER_P (value));
2753 shift = GET_MODE_BITSIZE (GET_MODE (value)) - GET_MODE_BITSIZE (mode);
2754 if (shift == 0)
2755 return false;
2757 /* Use ashr rather than lshr for right shifts. This is for the benefit
2758 of the MIPS port, which requires SImode values to be sign-extended
2759 when stored in 64-bit registers. */
2760 if (!force_expand_binop (GET_MODE (value), left_p ? ashl_optab : ashr_optab,
2761 value, GEN_INT (shift), value, 1, OPTAB_WIDEN))
2762 gcc_unreachable ();
2763 return true;
2766 /* If X is a likely-spilled register value, copy it to a pseudo
2767 register and return that register. Return X otherwise. */
2769 static rtx
2770 avoid_likely_spilled_reg (rtx x)
2772 rtx new_rtx;
2774 if (REG_P (x)
2775 && HARD_REGISTER_P (x)
2776 && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (x))))
2778 /* Make sure that we generate a REG rather than a CONCAT.
2779 Moves into CONCATs can need nontrivial instructions,
2780 and the whole point of this function is to avoid
2781 using the hard register directly in such a situation. */
2782 generating_concat_p = 0;
2783 new_rtx = gen_reg_rtx (GET_MODE (x));
2784 generating_concat_p = 1;
2785 emit_move_insn (new_rtx, x);
2786 return new_rtx;
2788 return x;
2791 /* Helper function for expand_call.
2792 Return false is EXP is not implementable as a sibling call. */
2794 static bool
2795 can_implement_as_sibling_call_p (tree exp,
2796 rtx structure_value_addr,
2797 tree funtype,
2798 int reg_parm_stack_space ATTRIBUTE_UNUSED,
2799 tree fndecl,
2800 int flags,
2801 tree addr,
2802 const args_size &args_size)
2804 if (!targetm.have_sibcall_epilogue ())
2806 maybe_complain_about_tail_call
2807 (exp,
2808 "machine description does not have"
2809 " a sibcall_epilogue instruction pattern");
2810 return false;
2813 /* Doing sibling call optimization needs some work, since
2814 structure_value_addr can be allocated on the stack.
2815 It does not seem worth the effort since few optimizable
2816 sibling calls will return a structure. */
2817 if (structure_value_addr != NULL_RTX)
2819 maybe_complain_about_tail_call (exp, "callee returns a structure");
2820 return false;
2823 #ifdef REG_PARM_STACK_SPACE
2824 /* If outgoing reg parm stack space changes, we can not do sibcall. */
2825 if (OUTGOING_REG_PARM_STACK_SPACE (funtype)
2826 != OUTGOING_REG_PARM_STACK_SPACE (TREE_TYPE (current_function_decl))
2827 || (reg_parm_stack_space != REG_PARM_STACK_SPACE (current_function_decl)))
2829 maybe_complain_about_tail_call (exp,
2830 "inconsistent size of stack space"
2831 " allocated for arguments which are"
2832 " passed in registers");
2833 return false;
2835 #endif
2837 /* Check whether the target is able to optimize the call
2838 into a sibcall. */
2839 if (!targetm.function_ok_for_sibcall (fndecl, exp))
2841 maybe_complain_about_tail_call (exp,
2842 "target is not able to optimize the"
2843 " call into a sibling call");
2844 return false;
2847 /* Functions that do not return exactly once may not be sibcall
2848 optimized. */
2849 if (flags & ECF_RETURNS_TWICE)
2851 maybe_complain_about_tail_call (exp, "callee returns twice");
2852 return false;
2854 if (flags & ECF_NORETURN)
2856 maybe_complain_about_tail_call (exp, "callee does not return");
2857 return false;
2860 if (TYPE_VOLATILE (TREE_TYPE (TREE_TYPE (addr))))
2862 maybe_complain_about_tail_call (exp, "volatile function type");
2863 return false;
2866 /* If the called function is nested in the current one, it might access
2867 some of the caller's arguments, but could clobber them beforehand if
2868 the argument areas are shared. */
2869 if (fndecl && decl_function_context (fndecl) == current_function_decl)
2871 maybe_complain_about_tail_call (exp, "nested function");
2872 return false;
2875 /* If this function requires more stack slots than the current
2876 function, we cannot change it into a sibling call.
2877 crtl->args.pretend_args_size is not part of the
2878 stack allocated by our caller. */
2879 if (args_size.constant > (crtl->args.size - crtl->args.pretend_args_size))
2881 maybe_complain_about_tail_call (exp,
2882 "callee required more stack slots"
2883 " than the caller");
2884 return false;
2887 /* If the callee pops its own arguments, then it must pop exactly
2888 the same number of arguments as the current function. */
2889 if (targetm.calls.return_pops_args (fndecl, funtype, args_size.constant)
2890 != targetm.calls.return_pops_args (current_function_decl,
2891 TREE_TYPE (current_function_decl),
2892 crtl->args.size))
2894 maybe_complain_about_tail_call (exp,
2895 "inconsistent number of"
2896 " popped arguments");
2897 return false;
2900 if (!lang_hooks.decls.ok_for_sibcall (fndecl))
2902 maybe_complain_about_tail_call (exp, "frontend does not support"
2903 " sibling call");
2904 return false;
2907 /* All checks passed. */
2908 return true;
2911 /* Generate all the code for a CALL_EXPR exp
2912 and return an rtx for its value.
2913 Store the value in TARGET (specified as an rtx) if convenient.
2914 If the value is stored in TARGET then TARGET is returned.
2915 If IGNORE is nonzero, then we ignore the value of the function call. */
2918 expand_call (tree exp, rtx target, int ignore)
2920 /* Nonzero if we are currently expanding a call. */
2921 static int currently_expanding_call = 0;
2923 /* RTX for the function to be called. */
2924 rtx funexp;
2925 /* Sequence of insns to perform a normal "call". */
2926 rtx_insn *normal_call_insns = NULL;
2927 /* Sequence of insns to perform a tail "call". */
2928 rtx_insn *tail_call_insns = NULL;
2929 /* Data type of the function. */
2930 tree funtype;
2931 tree type_arg_types;
2932 tree rettype;
2933 /* Declaration of the function being called,
2934 or 0 if the function is computed (not known by name). */
2935 tree fndecl = 0;
2936 /* The type of the function being called. */
2937 tree fntype;
2938 bool try_tail_call = CALL_EXPR_TAILCALL (exp);
2939 bool must_tail_call = CALL_EXPR_MUST_TAIL_CALL (exp);
2940 int pass;
2942 /* Register in which non-BLKmode value will be returned,
2943 or 0 if no value or if value is BLKmode. */
2944 rtx valreg;
2945 /* Register(s) in which bounds are returned. */
2946 rtx valbnd = NULL;
2947 /* Address where we should return a BLKmode value;
2948 0 if value not BLKmode. */
2949 rtx structure_value_addr = 0;
2950 /* Nonzero if that address is being passed by treating it as
2951 an extra, implicit first parameter. Otherwise,
2952 it is passed by being copied directly into struct_value_rtx. */
2953 int structure_value_addr_parm = 0;
2954 /* Holds the value of implicit argument for the struct value. */
2955 tree structure_value_addr_value = NULL_TREE;
2956 /* Size of aggregate value wanted, or zero if none wanted
2957 or if we are using the non-reentrant PCC calling convention
2958 or expecting the value in registers. */
2959 HOST_WIDE_INT struct_value_size = 0;
2960 /* Nonzero if called function returns an aggregate in memory PCC style,
2961 by returning the address of where to find it. */
2962 int pcc_struct_value = 0;
2963 rtx struct_value = 0;
2965 /* Number of actual parameters in this call, including struct value addr. */
2966 int num_actuals;
2967 /* Number of named args. Args after this are anonymous ones
2968 and they must all go on the stack. */
2969 int n_named_args;
2970 /* Number of complex actual arguments that need to be split. */
2971 int num_complex_actuals = 0;
2973 /* Vector of information about each argument.
2974 Arguments are numbered in the order they will be pushed,
2975 not the order they are written. */
2976 struct arg_data *args;
2978 /* Total size in bytes of all the stack-parms scanned so far. */
2979 struct args_size args_size;
2980 struct args_size adjusted_args_size;
2981 /* Size of arguments before any adjustments (such as rounding). */
2982 int unadjusted_args_size;
2983 /* Data on reg parms scanned so far. */
2984 CUMULATIVE_ARGS args_so_far_v;
2985 cumulative_args_t args_so_far;
2986 /* Nonzero if a reg parm has been scanned. */
2987 int reg_parm_seen;
2988 /* Nonzero if this is an indirect function call. */
2990 /* Nonzero if we must avoid push-insns in the args for this call.
2991 If stack space is allocated for register parameters, but not by the
2992 caller, then it is preallocated in the fixed part of the stack frame.
2993 So the entire argument block must then be preallocated (i.e., we
2994 ignore PUSH_ROUNDING in that case). */
2996 int must_preallocate = !PUSH_ARGS;
2998 /* Size of the stack reserved for parameter registers. */
2999 int reg_parm_stack_space = 0;
3001 /* Address of space preallocated for stack parms
3002 (on machines that lack push insns), or 0 if space not preallocated. */
3003 rtx argblock = 0;
3005 /* Mask of ECF_ and ERF_ flags. */
3006 int flags = 0;
3007 int return_flags = 0;
3008 #ifdef REG_PARM_STACK_SPACE
3009 /* Define the boundary of the register parm stack space that needs to be
3010 saved, if any. */
3011 int low_to_save, high_to_save;
3012 rtx save_area = 0; /* Place that it is saved */
3013 #endif
3015 int initial_highest_arg_in_use = highest_outgoing_arg_in_use;
3016 char *initial_stack_usage_map = stack_usage_map;
3017 char *stack_usage_map_buf = NULL;
3019 int old_stack_allocated;
3021 /* State variables to track stack modifications. */
3022 rtx old_stack_level = 0;
3023 int old_stack_arg_under_construction = 0;
3024 int old_pending_adj = 0;
3025 int old_inhibit_defer_pop = inhibit_defer_pop;
3027 /* Some stack pointer alterations we make are performed via
3028 allocate_dynamic_stack_space. This modifies the stack_pointer_delta,
3029 which we then also need to save/restore along the way. */
3030 int old_stack_pointer_delta = 0;
3032 rtx call_fusage;
3033 tree addr = CALL_EXPR_FN (exp);
3034 int i;
3035 /* The alignment of the stack, in bits. */
3036 unsigned HOST_WIDE_INT preferred_stack_boundary;
3037 /* The alignment of the stack, in bytes. */
3038 unsigned HOST_WIDE_INT preferred_unit_stack_boundary;
3039 /* The static chain value to use for this call. */
3040 rtx static_chain_value;
3041 /* See if this is "nothrow" function call. */
3042 if (TREE_NOTHROW (exp))
3043 flags |= ECF_NOTHROW;
3045 /* See if we can find a DECL-node for the actual function, and get the
3046 function attributes (flags) from the function decl or type node. */
3047 fndecl = get_callee_fndecl (exp);
3048 if (fndecl)
3050 fntype = TREE_TYPE (fndecl);
3051 flags |= flags_from_decl_or_type (fndecl);
3052 return_flags |= decl_return_flags (fndecl);
3054 else
3056 fntype = TREE_TYPE (TREE_TYPE (addr));
3057 flags |= flags_from_decl_or_type (fntype);
3058 if (CALL_EXPR_BY_DESCRIPTOR (exp))
3059 flags |= ECF_BY_DESCRIPTOR;
3061 rettype = TREE_TYPE (exp);
3063 struct_value = targetm.calls.struct_value_rtx (fntype, 0);
3065 /* Warn if this value is an aggregate type,
3066 regardless of which calling convention we are using for it. */
3067 if (AGGREGATE_TYPE_P (rettype))
3068 warning (OPT_Waggregate_return, "function call has aggregate value");
3070 /* If the result of a non looping pure or const function call is
3071 ignored (or void), and none of its arguments are volatile, we can
3072 avoid expanding the call and just evaluate the arguments for
3073 side-effects. */
3074 if ((flags & (ECF_CONST | ECF_PURE))
3075 && (!(flags & ECF_LOOPING_CONST_OR_PURE))
3076 && (ignore || target == const0_rtx
3077 || TYPE_MODE (rettype) == VOIDmode))
3079 bool volatilep = false;
3080 tree arg;
3081 call_expr_arg_iterator iter;
3083 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
3084 if (TREE_THIS_VOLATILE (arg))
3086 volatilep = true;
3087 break;
3090 if (! volatilep)
3092 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
3093 expand_expr (arg, const0_rtx, VOIDmode, EXPAND_NORMAL);
3094 return const0_rtx;
3098 #ifdef REG_PARM_STACK_SPACE
3099 reg_parm_stack_space = REG_PARM_STACK_SPACE (!fndecl ? fntype : fndecl);
3100 #endif
3102 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))
3103 && reg_parm_stack_space > 0 && PUSH_ARGS)
3104 must_preallocate = 1;
3106 /* Set up a place to return a structure. */
3108 /* Cater to broken compilers. */
3109 if (aggregate_value_p (exp, fntype))
3111 /* This call returns a big structure. */
3112 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
3114 #ifdef PCC_STATIC_STRUCT_RETURN
3116 pcc_struct_value = 1;
3118 #else /* not PCC_STATIC_STRUCT_RETURN */
3120 struct_value_size = int_size_in_bytes (rettype);
3122 /* Even if it is semantically safe to use the target as the return
3123 slot, it may be not sufficiently aligned for the return type. */
3124 if (CALL_EXPR_RETURN_SLOT_OPT (exp)
3125 && target
3126 && MEM_P (target)
3127 && !(MEM_ALIGN (target) < TYPE_ALIGN (rettype)
3128 && SLOW_UNALIGNED_ACCESS (TYPE_MODE (rettype),
3129 MEM_ALIGN (target))))
3130 structure_value_addr = XEXP (target, 0);
3131 else
3133 /* For variable-sized objects, we must be called with a target
3134 specified. If we were to allocate space on the stack here,
3135 we would have no way of knowing when to free it. */
3136 rtx d = assign_temp (rettype, 1, 1);
3137 structure_value_addr = XEXP (d, 0);
3138 target = 0;
3141 #endif /* not PCC_STATIC_STRUCT_RETURN */
3144 /* Figure out the amount to which the stack should be aligned. */
3145 preferred_stack_boundary = PREFERRED_STACK_BOUNDARY;
3146 if (fndecl)
3148 struct cgraph_rtl_info *i = cgraph_node::rtl_info (fndecl);
3149 /* Without automatic stack alignment, we can't increase preferred
3150 stack boundary. With automatic stack alignment, it is
3151 unnecessary since unless we can guarantee that all callers will
3152 align the outgoing stack properly, callee has to align its
3153 stack anyway. */
3154 if (i
3155 && i->preferred_incoming_stack_boundary
3156 && i->preferred_incoming_stack_boundary < preferred_stack_boundary)
3157 preferred_stack_boundary = i->preferred_incoming_stack_boundary;
3160 /* Operand 0 is a pointer-to-function; get the type of the function. */
3161 funtype = TREE_TYPE (addr);
3162 gcc_assert (POINTER_TYPE_P (funtype));
3163 funtype = TREE_TYPE (funtype);
3165 /* Count whether there are actual complex arguments that need to be split
3166 into their real and imaginary parts. Munge the type_arg_types
3167 appropriately here as well. */
3168 if (targetm.calls.split_complex_arg)
3170 call_expr_arg_iterator iter;
3171 tree arg;
3172 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
3174 tree type = TREE_TYPE (arg);
3175 if (type && TREE_CODE (type) == COMPLEX_TYPE
3176 && targetm.calls.split_complex_arg (type))
3177 num_complex_actuals++;
3179 type_arg_types = split_complex_types (TYPE_ARG_TYPES (funtype));
3181 else
3182 type_arg_types = TYPE_ARG_TYPES (funtype);
3184 if (flags & ECF_MAY_BE_ALLOCA)
3185 cfun->calls_alloca = 1;
3187 /* If struct_value_rtx is 0, it means pass the address
3188 as if it were an extra parameter. Put the argument expression
3189 in structure_value_addr_value. */
3190 if (structure_value_addr && struct_value == 0)
3192 /* If structure_value_addr is a REG other than
3193 virtual_outgoing_args_rtx, we can use always use it. If it
3194 is not a REG, we must always copy it into a register.
3195 If it is virtual_outgoing_args_rtx, we must copy it to another
3196 register in some cases. */
3197 rtx temp = (!REG_P (structure_value_addr)
3198 || (ACCUMULATE_OUTGOING_ARGS
3199 && stack_arg_under_construction
3200 && structure_value_addr == virtual_outgoing_args_rtx)
3201 ? copy_addr_to_reg (convert_memory_address
3202 (Pmode, structure_value_addr))
3203 : structure_value_addr);
3205 structure_value_addr_value =
3206 make_tree (build_pointer_type (TREE_TYPE (funtype)), temp);
3207 structure_value_addr_parm = CALL_WITH_BOUNDS_P (exp) ? 2 : 1;
3210 /* Count the arguments and set NUM_ACTUALS. */
3211 num_actuals =
3212 call_expr_nargs (exp) + num_complex_actuals + structure_value_addr_parm;
3214 /* Compute number of named args.
3215 First, do a raw count of the args for INIT_CUMULATIVE_ARGS. */
3217 if (type_arg_types != 0)
3218 n_named_args
3219 = (list_length (type_arg_types)
3220 /* Count the struct value address, if it is passed as a parm. */
3221 + structure_value_addr_parm);
3222 else
3223 /* If we know nothing, treat all args as named. */
3224 n_named_args = num_actuals;
3226 /* Start updating where the next arg would go.
3228 On some machines (such as the PA) indirect calls have a different
3229 calling convention than normal calls. The fourth argument in
3230 INIT_CUMULATIVE_ARGS tells the backend if this is an indirect call
3231 or not. */
3232 INIT_CUMULATIVE_ARGS (args_so_far_v, funtype, NULL_RTX, fndecl, n_named_args);
3233 args_so_far = pack_cumulative_args (&args_so_far_v);
3235 /* Now possibly adjust the number of named args.
3236 Normally, don't include the last named arg if anonymous args follow.
3237 We do include the last named arg if
3238 targetm.calls.strict_argument_naming() returns nonzero.
3239 (If no anonymous args follow, the result of list_length is actually
3240 one too large. This is harmless.)
3242 If targetm.calls.pretend_outgoing_varargs_named() returns
3243 nonzero, and targetm.calls.strict_argument_naming() returns zero,
3244 this machine will be able to place unnamed args that were passed
3245 in registers into the stack. So treat all args as named. This
3246 allows the insns emitting for a specific argument list to be
3247 independent of the function declaration.
3249 If targetm.calls.pretend_outgoing_varargs_named() returns zero,
3250 we do not have any reliable way to pass unnamed args in
3251 registers, so we must force them into memory. */
3253 if (type_arg_types != 0
3254 && targetm.calls.strict_argument_naming (args_so_far))
3256 else if (type_arg_types != 0
3257 && ! targetm.calls.pretend_outgoing_varargs_named (args_so_far))
3258 /* Don't include the last named arg. */
3259 --n_named_args;
3260 else
3261 /* Treat all args as named. */
3262 n_named_args = num_actuals;
3264 /* Make a vector to hold all the information about each arg. */
3265 args = XCNEWVEC (struct arg_data, num_actuals);
3267 /* Build up entries in the ARGS array, compute the size of the
3268 arguments into ARGS_SIZE, etc. */
3269 initialize_argument_information (num_actuals, args, &args_size,
3270 n_named_args, exp,
3271 structure_value_addr_value, fndecl, fntype,
3272 args_so_far, reg_parm_stack_space,
3273 &old_stack_level, &old_pending_adj,
3274 &must_preallocate, &flags,
3275 &try_tail_call, CALL_FROM_THUNK_P (exp));
3277 if (args_size.var)
3278 must_preallocate = 1;
3280 /* Now make final decision about preallocating stack space. */
3281 must_preallocate = finalize_must_preallocate (must_preallocate,
3282 num_actuals, args,
3283 &args_size);
3285 /* If the structure value address will reference the stack pointer, we
3286 must stabilize it. We don't need to do this if we know that we are
3287 not going to adjust the stack pointer in processing this call. */
3289 if (structure_value_addr
3290 && (reg_mentioned_p (virtual_stack_dynamic_rtx, structure_value_addr)
3291 || reg_mentioned_p (virtual_outgoing_args_rtx,
3292 structure_value_addr))
3293 && (args_size.var
3294 || (!ACCUMULATE_OUTGOING_ARGS && args_size.constant)))
3295 structure_value_addr = copy_to_reg (structure_value_addr);
3297 /* Tail calls can make things harder to debug, and we've traditionally
3298 pushed these optimizations into -O2. Don't try if we're already
3299 expanding a call, as that means we're an argument. Don't try if
3300 there's cleanups, as we know there's code to follow the call. */
3302 if (currently_expanding_call++ != 0
3303 || !flag_optimize_sibling_calls
3304 || args_size.var
3305 || dbg_cnt (tail_call) == false)
3306 try_tail_call = 0;
3308 /* If the user has marked the function as requiring tail-call
3309 optimization, attempt it. */
3310 if (must_tail_call)
3311 try_tail_call = 1;
3313 /* Rest of purposes for tail call optimizations to fail. */
3314 if (try_tail_call)
3315 try_tail_call = can_implement_as_sibling_call_p (exp,
3316 structure_value_addr,
3317 funtype,
3318 reg_parm_stack_space,
3319 fndecl,
3320 flags, addr, args_size);
3322 /* Check if caller and callee disagree in promotion of function
3323 return value. */
3324 if (try_tail_call)
3326 machine_mode caller_mode, caller_promoted_mode;
3327 machine_mode callee_mode, callee_promoted_mode;
3328 int caller_unsignedp, callee_unsignedp;
3329 tree caller_res = DECL_RESULT (current_function_decl);
3331 caller_unsignedp = TYPE_UNSIGNED (TREE_TYPE (caller_res));
3332 caller_mode = DECL_MODE (caller_res);
3333 callee_unsignedp = TYPE_UNSIGNED (TREE_TYPE (funtype));
3334 callee_mode = TYPE_MODE (TREE_TYPE (funtype));
3335 caller_promoted_mode
3336 = promote_function_mode (TREE_TYPE (caller_res), caller_mode,
3337 &caller_unsignedp,
3338 TREE_TYPE (current_function_decl), 1);
3339 callee_promoted_mode
3340 = promote_function_mode (TREE_TYPE (funtype), callee_mode,
3341 &callee_unsignedp,
3342 funtype, 1);
3343 if (caller_mode != VOIDmode
3344 && (caller_promoted_mode != callee_promoted_mode
3345 || ((caller_mode != caller_promoted_mode
3346 || callee_mode != callee_promoted_mode)
3347 && (caller_unsignedp != callee_unsignedp
3348 || GET_MODE_BITSIZE (caller_mode)
3349 < GET_MODE_BITSIZE (callee_mode)))))
3351 try_tail_call = 0;
3352 maybe_complain_about_tail_call (exp,
3353 "caller and callee disagree in"
3354 " promotion of function"
3355 " return value");
3359 /* Ensure current function's preferred stack boundary is at least
3360 what we need. Stack alignment may also increase preferred stack
3361 boundary. */
3362 if (crtl->preferred_stack_boundary < preferred_stack_boundary)
3363 crtl->preferred_stack_boundary = preferred_stack_boundary;
3364 else
3365 preferred_stack_boundary = crtl->preferred_stack_boundary;
3367 preferred_unit_stack_boundary = preferred_stack_boundary / BITS_PER_UNIT;
3369 /* We want to make two insn chains; one for a sibling call, the other
3370 for a normal call. We will select one of the two chains after
3371 initial RTL generation is complete. */
3372 for (pass = try_tail_call ? 0 : 1; pass < 2; pass++)
3374 int sibcall_failure = 0;
3375 /* We want to emit any pending stack adjustments before the tail
3376 recursion "call". That way we know any adjustment after the tail
3377 recursion call can be ignored if we indeed use the tail
3378 call expansion. */
3379 saved_pending_stack_adjust save;
3380 rtx_insn *insns, *before_call, *after_args;
3381 rtx next_arg_reg;
3383 if (pass == 0)
3385 /* State variables we need to save and restore between
3386 iterations. */
3387 save_pending_stack_adjust (&save);
3389 if (pass)
3390 flags &= ~ECF_SIBCALL;
3391 else
3392 flags |= ECF_SIBCALL;
3394 /* Other state variables that we must reinitialize each time
3395 through the loop (that are not initialized by the loop itself). */
3396 argblock = 0;
3397 call_fusage = 0;
3399 /* Start a new sequence for the normal call case.
3401 From this point on, if the sibling call fails, we want to set
3402 sibcall_failure instead of continuing the loop. */
3403 start_sequence ();
3405 /* Don't let pending stack adjusts add up to too much.
3406 Also, do all pending adjustments now if there is any chance
3407 this might be a call to alloca or if we are expanding a sibling
3408 call sequence.
3409 Also do the adjustments before a throwing call, otherwise
3410 exception handling can fail; PR 19225. */
3411 if (pending_stack_adjust >= 32
3412 || (pending_stack_adjust > 0
3413 && (flags & ECF_MAY_BE_ALLOCA))
3414 || (pending_stack_adjust > 0
3415 && flag_exceptions && !(flags & ECF_NOTHROW))
3416 || pass == 0)
3417 do_pending_stack_adjust ();
3419 /* Precompute any arguments as needed. */
3420 if (pass)
3421 precompute_arguments (num_actuals, args);
3423 /* Now we are about to start emitting insns that can be deleted
3424 if a libcall is deleted. */
3425 if (pass && (flags & ECF_MALLOC))
3426 start_sequence ();
3428 if (pass == 0
3429 && crtl->stack_protect_guard
3430 && targetm.stack_protect_runtime_enabled_p ())
3431 stack_protect_epilogue ();
3433 adjusted_args_size = args_size;
3434 /* Compute the actual size of the argument block required. The variable
3435 and constant sizes must be combined, the size may have to be rounded,
3436 and there may be a minimum required size. When generating a sibcall
3437 pattern, do not round up, since we'll be re-using whatever space our
3438 caller provided. */
3439 unadjusted_args_size
3440 = compute_argument_block_size (reg_parm_stack_space,
3441 &adjusted_args_size,
3442 fndecl, fntype,
3443 (pass == 0 ? 0
3444 : preferred_stack_boundary));
3446 old_stack_allocated = stack_pointer_delta - pending_stack_adjust;
3448 /* The argument block when performing a sibling call is the
3449 incoming argument block. */
3450 if (pass == 0)
3452 argblock = crtl->args.internal_arg_pointer;
3453 if (STACK_GROWS_DOWNWARD)
3454 argblock
3455 = plus_constant (Pmode, argblock, crtl->args.pretend_args_size);
3456 else
3457 argblock
3458 = plus_constant (Pmode, argblock, -crtl->args.pretend_args_size);
3460 stored_args_map = sbitmap_alloc (args_size.constant);
3461 bitmap_clear (stored_args_map);
3464 /* If we have no actual push instructions, or shouldn't use them,
3465 make space for all args right now. */
3466 else if (adjusted_args_size.var != 0)
3468 if (old_stack_level == 0)
3470 emit_stack_save (SAVE_BLOCK, &old_stack_level);
3471 old_stack_pointer_delta = stack_pointer_delta;
3472 old_pending_adj = pending_stack_adjust;
3473 pending_stack_adjust = 0;
3474 /* stack_arg_under_construction says whether a stack arg is
3475 being constructed at the old stack level. Pushing the stack
3476 gets a clean outgoing argument block. */
3477 old_stack_arg_under_construction = stack_arg_under_construction;
3478 stack_arg_under_construction = 0;
3480 argblock = push_block (ARGS_SIZE_RTX (adjusted_args_size), 0, 0);
3481 if (flag_stack_usage_info)
3482 current_function_has_unbounded_dynamic_stack_size = 1;
3484 else
3486 /* Note that we must go through the motions of allocating an argument
3487 block even if the size is zero because we may be storing args
3488 in the area reserved for register arguments, which may be part of
3489 the stack frame. */
3491 int needed = adjusted_args_size.constant;
3493 /* Store the maximum argument space used. It will be pushed by
3494 the prologue (if ACCUMULATE_OUTGOING_ARGS, or stack overflow
3495 checking). */
3497 if (needed > crtl->outgoing_args_size)
3498 crtl->outgoing_args_size = needed;
3500 if (must_preallocate)
3502 if (ACCUMULATE_OUTGOING_ARGS)
3504 /* Since the stack pointer will never be pushed, it is
3505 possible for the evaluation of a parm to clobber
3506 something we have already written to the stack.
3507 Since most function calls on RISC machines do not use
3508 the stack, this is uncommon, but must work correctly.
3510 Therefore, we save any area of the stack that was already
3511 written and that we are using. Here we set up to do this
3512 by making a new stack usage map from the old one. The
3513 actual save will be done by store_one_arg.
3515 Another approach might be to try to reorder the argument
3516 evaluations to avoid this conflicting stack usage. */
3518 /* Since we will be writing into the entire argument area,
3519 the map must be allocated for its entire size, not just
3520 the part that is the responsibility of the caller. */
3521 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
3522 needed += reg_parm_stack_space;
3524 if (ARGS_GROW_DOWNWARD)
3525 highest_outgoing_arg_in_use
3526 = MAX (initial_highest_arg_in_use, needed + 1);
3527 else
3528 highest_outgoing_arg_in_use
3529 = MAX (initial_highest_arg_in_use, needed);
3531 free (stack_usage_map_buf);
3532 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use);
3533 stack_usage_map = stack_usage_map_buf;
3535 if (initial_highest_arg_in_use)
3536 memcpy (stack_usage_map, initial_stack_usage_map,
3537 initial_highest_arg_in_use);
3539 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use)
3540 memset (&stack_usage_map[initial_highest_arg_in_use], 0,
3541 (highest_outgoing_arg_in_use
3542 - initial_highest_arg_in_use));
3543 needed = 0;
3545 /* The address of the outgoing argument list must not be
3546 copied to a register here, because argblock would be left
3547 pointing to the wrong place after the call to
3548 allocate_dynamic_stack_space below. */
3550 argblock = virtual_outgoing_args_rtx;
3552 else
3554 if (inhibit_defer_pop == 0)
3556 /* Try to reuse some or all of the pending_stack_adjust
3557 to get this space. */
3558 needed
3559 = (combine_pending_stack_adjustment_and_call
3560 (unadjusted_args_size,
3561 &adjusted_args_size,
3562 preferred_unit_stack_boundary));
3564 /* combine_pending_stack_adjustment_and_call computes
3565 an adjustment before the arguments are allocated.
3566 Account for them and see whether or not the stack
3567 needs to go up or down. */
3568 needed = unadjusted_args_size - needed;
3570 if (needed < 0)
3572 /* We're releasing stack space. */
3573 /* ??? We can avoid any adjustment at all if we're
3574 already aligned. FIXME. */
3575 pending_stack_adjust = -needed;
3576 do_pending_stack_adjust ();
3577 needed = 0;
3579 else
3580 /* We need to allocate space. We'll do that in
3581 push_block below. */
3582 pending_stack_adjust = 0;
3585 /* Special case this because overhead of `push_block' in
3586 this case is non-trivial. */
3587 if (needed == 0)
3588 argblock = virtual_outgoing_args_rtx;
3589 else
3591 argblock = push_block (GEN_INT (needed), 0, 0);
3592 if (ARGS_GROW_DOWNWARD)
3593 argblock = plus_constant (Pmode, argblock, needed);
3596 /* We only really need to call `copy_to_reg' in the case
3597 where push insns are going to be used to pass ARGBLOCK
3598 to a function call in ARGS. In that case, the stack
3599 pointer changes value from the allocation point to the
3600 call point, and hence the value of
3601 VIRTUAL_OUTGOING_ARGS_RTX changes as well. But might
3602 as well always do it. */
3603 argblock = copy_to_reg (argblock);
3608 if (ACCUMULATE_OUTGOING_ARGS)
3610 /* The save/restore code in store_one_arg handles all
3611 cases except one: a constructor call (including a C
3612 function returning a BLKmode struct) to initialize
3613 an argument. */
3614 if (stack_arg_under_construction)
3616 rtx push_size
3617 = GEN_INT (adjusted_args_size.constant
3618 + (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype
3619 : TREE_TYPE (fndecl))) ? 0
3620 : reg_parm_stack_space));
3621 if (old_stack_level == 0)
3623 emit_stack_save (SAVE_BLOCK, &old_stack_level);
3624 old_stack_pointer_delta = stack_pointer_delta;
3625 old_pending_adj = pending_stack_adjust;
3626 pending_stack_adjust = 0;
3627 /* stack_arg_under_construction says whether a stack
3628 arg is being constructed at the old stack level.
3629 Pushing the stack gets a clean outgoing argument
3630 block. */
3631 old_stack_arg_under_construction
3632 = stack_arg_under_construction;
3633 stack_arg_under_construction = 0;
3634 /* Make a new map for the new argument list. */
3635 free (stack_usage_map_buf);
3636 stack_usage_map_buf = XCNEWVEC (char, highest_outgoing_arg_in_use);
3637 stack_usage_map = stack_usage_map_buf;
3638 highest_outgoing_arg_in_use = 0;
3640 /* We can pass TRUE as the 4th argument because we just
3641 saved the stack pointer and will restore it right after
3642 the call. */
3643 allocate_dynamic_stack_space (push_size, 0,
3644 BIGGEST_ALIGNMENT, true);
3647 /* If argument evaluation might modify the stack pointer,
3648 copy the address of the argument list to a register. */
3649 for (i = 0; i < num_actuals; i++)
3650 if (args[i].pass_on_stack)
3652 argblock = copy_addr_to_reg (argblock);
3653 break;
3657 compute_argument_addresses (args, argblock, num_actuals);
3659 /* Stack is properly aligned, pops can't safely be deferred during
3660 the evaluation of the arguments. */
3661 NO_DEFER_POP;
3663 /* Precompute all register parameters. It isn't safe to compute
3664 anything once we have started filling any specific hard regs.
3665 TLS symbols sometimes need a call to resolve. Precompute
3666 register parameters before any stack pointer manipulation
3667 to avoid unaligned stack in the called function. */
3668 precompute_register_parameters (num_actuals, args, &reg_parm_seen);
3670 OK_DEFER_POP;
3672 /* Perform stack alignment before the first push (the last arg). */
3673 if (argblock == 0
3674 && adjusted_args_size.constant > reg_parm_stack_space
3675 && adjusted_args_size.constant != unadjusted_args_size)
3677 /* When the stack adjustment is pending, we get better code
3678 by combining the adjustments. */
3679 if (pending_stack_adjust
3680 && ! inhibit_defer_pop)
3682 pending_stack_adjust
3683 = (combine_pending_stack_adjustment_and_call
3684 (unadjusted_args_size,
3685 &adjusted_args_size,
3686 preferred_unit_stack_boundary));
3687 do_pending_stack_adjust ();
3689 else if (argblock == 0)
3690 anti_adjust_stack (GEN_INT (adjusted_args_size.constant
3691 - unadjusted_args_size));
3693 /* Now that the stack is properly aligned, pops can't safely
3694 be deferred during the evaluation of the arguments. */
3695 NO_DEFER_POP;
3697 /* Record the maximum pushed stack space size. We need to delay
3698 doing it this far to take into account the optimization done
3699 by combine_pending_stack_adjustment_and_call. */
3700 if (flag_stack_usage_info
3701 && !ACCUMULATE_OUTGOING_ARGS
3702 && pass
3703 && adjusted_args_size.var == 0)
3705 int pushed = adjusted_args_size.constant + pending_stack_adjust;
3706 if (pushed > current_function_pushed_stack_size)
3707 current_function_pushed_stack_size = pushed;
3710 funexp = rtx_for_function_call (fndecl, addr);
3712 if (CALL_EXPR_STATIC_CHAIN (exp))
3713 static_chain_value = expand_normal (CALL_EXPR_STATIC_CHAIN (exp));
3714 else
3715 static_chain_value = 0;
3717 #ifdef REG_PARM_STACK_SPACE
3718 /* Save the fixed argument area if it's part of the caller's frame and
3719 is clobbered by argument setup for this call. */
3720 if (ACCUMULATE_OUTGOING_ARGS && pass)
3721 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock,
3722 &low_to_save, &high_to_save);
3723 #endif
3725 /* Now store (and compute if necessary) all non-register parms.
3726 These come before register parms, since they can require block-moves,
3727 which could clobber the registers used for register parms.
3728 Parms which have partial registers are not stored here,
3729 but we do preallocate space here if they want that. */
3731 for (i = 0; i < num_actuals; i++)
3733 /* Delay bounds until all other args are stored. */
3734 if (POINTER_BOUNDS_P (args[i].tree_value))
3735 continue;
3736 else if (args[i].reg == 0 || args[i].pass_on_stack)
3738 rtx_insn *before_arg = get_last_insn ();
3740 /* We don't allow passing huge (> 2^30 B) arguments
3741 by value. It would cause an overflow later on. */
3742 if (adjusted_args_size.constant
3743 >= (1 << (HOST_BITS_PER_INT - 2)))
3745 sorry ("passing too large argument on stack");
3746 continue;
3749 if (store_one_arg (&args[i], argblock, flags,
3750 adjusted_args_size.var != 0,
3751 reg_parm_stack_space)
3752 || (pass == 0
3753 && check_sibcall_argument_overlap (before_arg,
3754 &args[i], 1)))
3755 sibcall_failure = 1;
3758 if (args[i].stack)
3759 call_fusage
3760 = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[i].tree_value)),
3761 gen_rtx_USE (VOIDmode, args[i].stack),
3762 call_fusage);
3765 /* If we have a parm that is passed in registers but not in memory
3766 and whose alignment does not permit a direct copy into registers,
3767 make a group of pseudos that correspond to each register that we
3768 will later fill. */
3769 if (STRICT_ALIGNMENT)
3770 store_unaligned_arguments_into_pseudos (args, num_actuals);
3772 /* Now store any partially-in-registers parm.
3773 This is the last place a block-move can happen. */
3774 if (reg_parm_seen)
3775 for (i = 0; i < num_actuals; i++)
3776 if (args[i].partial != 0 && ! args[i].pass_on_stack)
3778 rtx_insn *before_arg = get_last_insn ();
3780 /* On targets with weird calling conventions (e.g. PA) it's
3781 hard to ensure that all cases of argument overlap between
3782 stack and registers work. Play it safe and bail out. */
3783 if (ARGS_GROW_DOWNWARD && !STACK_GROWS_DOWNWARD)
3785 sibcall_failure = 1;
3786 break;
3789 if (store_one_arg (&args[i], argblock, flags,
3790 adjusted_args_size.var != 0,
3791 reg_parm_stack_space)
3792 || (pass == 0
3793 && check_sibcall_argument_overlap (before_arg,
3794 &args[i], 1)))
3795 sibcall_failure = 1;
3798 bool any_regs = false;
3799 for (i = 0; i < num_actuals; i++)
3800 if (args[i].reg != NULL_RTX)
3802 any_regs = true;
3803 targetm.calls.call_args (args[i].reg, funtype);
3805 if (!any_regs)
3806 targetm.calls.call_args (pc_rtx, funtype);
3808 /* Figure out the register where the value, if any, will come back. */
3809 valreg = 0;
3810 valbnd = 0;
3811 if (TYPE_MODE (rettype) != VOIDmode
3812 && ! structure_value_addr)
3814 if (pcc_struct_value)
3816 valreg = hard_function_value (build_pointer_type (rettype),
3817 fndecl, NULL, (pass == 0));
3818 if (CALL_WITH_BOUNDS_P (exp))
3819 valbnd = targetm.calls.
3820 chkp_function_value_bounds (build_pointer_type (rettype),
3821 fndecl, (pass == 0));
3823 else
3825 valreg = hard_function_value (rettype, fndecl, fntype,
3826 (pass == 0));
3827 if (CALL_WITH_BOUNDS_P (exp))
3828 valbnd = targetm.calls.chkp_function_value_bounds (rettype,
3829 fndecl,
3830 (pass == 0));
3833 /* If VALREG is a PARALLEL whose first member has a zero
3834 offset, use that. This is for targets such as m68k that
3835 return the same value in multiple places. */
3836 if (GET_CODE (valreg) == PARALLEL)
3838 rtx elem = XVECEXP (valreg, 0, 0);
3839 rtx where = XEXP (elem, 0);
3840 rtx offset = XEXP (elem, 1);
3841 if (offset == const0_rtx
3842 && GET_MODE (where) == GET_MODE (valreg))
3843 valreg = where;
3847 /* Store all bounds not passed in registers. */
3848 for (i = 0; i < num_actuals; i++)
3850 if (POINTER_BOUNDS_P (args[i].tree_value)
3851 && !args[i].reg)
3852 store_bounds (&args[i],
3853 args[i].pointer_arg == -1
3854 ? NULL
3855 : &args[args[i].pointer_arg]);
3858 /* If register arguments require space on the stack and stack space
3859 was not preallocated, allocate stack space here for arguments
3860 passed in registers. */
3861 if (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))
3862 && !ACCUMULATE_OUTGOING_ARGS
3863 && must_preallocate == 0 && reg_parm_stack_space > 0)
3864 anti_adjust_stack (GEN_INT (reg_parm_stack_space));
3866 /* Pass the function the address in which to return a
3867 structure value. */
3868 if (pass != 0 && structure_value_addr && ! structure_value_addr_parm)
3870 structure_value_addr
3871 = convert_memory_address (Pmode, structure_value_addr);
3872 emit_move_insn (struct_value,
3873 force_reg (Pmode,
3874 force_operand (structure_value_addr,
3875 NULL_RTX)));
3877 if (REG_P (struct_value))
3878 use_reg (&call_fusage, struct_value);
3881 after_args = get_last_insn ();
3882 funexp = prepare_call_address (fndecl ? fndecl : fntype, funexp,
3883 static_chain_value, &call_fusage,
3884 reg_parm_seen, flags);
3886 load_register_parameters (args, num_actuals, &call_fusage, flags,
3887 pass == 0, &sibcall_failure);
3889 /* Save a pointer to the last insn before the call, so that we can
3890 later safely search backwards to find the CALL_INSN. */
3891 before_call = get_last_insn ();
3893 /* Set up next argument register. For sibling calls on machines
3894 with register windows this should be the incoming register. */
3895 if (pass == 0)
3896 next_arg_reg = targetm.calls.function_incoming_arg (args_so_far,
3897 VOIDmode,
3898 void_type_node,
3899 true);
3900 else
3901 next_arg_reg = targetm.calls.function_arg (args_so_far,
3902 VOIDmode, void_type_node,
3903 true);
3905 if (pass == 1 && (return_flags & ERF_RETURNS_ARG))
3907 int arg_nr = return_flags & ERF_RETURN_ARG_MASK;
3908 arg_nr = num_actuals - arg_nr - 1;
3909 if (arg_nr >= 0
3910 && arg_nr < num_actuals
3911 && args[arg_nr].reg
3912 && valreg
3913 && REG_P (valreg)
3914 && GET_MODE (args[arg_nr].reg) == GET_MODE (valreg))
3915 call_fusage
3916 = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[arg_nr].tree_value)),
3917 gen_rtx_SET (valreg, args[arg_nr].reg),
3918 call_fusage);
3920 /* All arguments and registers used for the call must be set up by
3921 now! */
3923 /* Stack must be properly aligned now. */
3924 gcc_assert (!pass
3925 || !(stack_pointer_delta % preferred_unit_stack_boundary));
3927 /* Generate the actual call instruction. */
3928 emit_call_1 (funexp, exp, fndecl, funtype, unadjusted_args_size,
3929 adjusted_args_size.constant, struct_value_size,
3930 next_arg_reg, valreg, old_inhibit_defer_pop, call_fusage,
3931 flags, args_so_far);
3933 if (flag_ipa_ra)
3935 rtx_call_insn *last;
3936 rtx datum = NULL_RTX;
3937 if (fndecl != NULL_TREE)
3939 datum = XEXP (DECL_RTL (fndecl), 0);
3940 gcc_assert (datum != NULL_RTX
3941 && GET_CODE (datum) == SYMBOL_REF);
3943 last = last_call_insn ();
3944 add_reg_note (last, REG_CALL_DECL, datum);
3947 /* If the call setup or the call itself overlaps with anything
3948 of the argument setup we probably clobbered our call address.
3949 In that case we can't do sibcalls. */
3950 if (pass == 0
3951 && check_sibcall_argument_overlap (after_args, 0, 0))
3952 sibcall_failure = 1;
3954 /* If a non-BLKmode value is returned at the most significant end
3955 of a register, shift the register right by the appropriate amount
3956 and update VALREG accordingly. BLKmode values are handled by the
3957 group load/store machinery below. */
3958 if (!structure_value_addr
3959 && !pcc_struct_value
3960 && TYPE_MODE (rettype) != VOIDmode
3961 && TYPE_MODE (rettype) != BLKmode
3962 && REG_P (valreg)
3963 && targetm.calls.return_in_msb (rettype))
3965 if (shift_return_value (TYPE_MODE (rettype), false, valreg))
3966 sibcall_failure = 1;
3967 valreg = gen_rtx_REG (TYPE_MODE (rettype), REGNO (valreg));
3970 if (pass && (flags & ECF_MALLOC))
3972 rtx temp = gen_reg_rtx (GET_MODE (valreg));
3973 rtx_insn *last, *insns;
3975 /* The return value from a malloc-like function is a pointer. */
3976 if (TREE_CODE (rettype) == POINTER_TYPE)
3977 mark_reg_pointer (temp, MALLOC_ABI_ALIGNMENT);
3979 emit_move_insn (temp, valreg);
3981 /* The return value from a malloc-like function can not alias
3982 anything else. */
3983 last = get_last_insn ();
3984 add_reg_note (last, REG_NOALIAS, temp);
3986 /* Write out the sequence. */
3987 insns = get_insns ();
3988 end_sequence ();
3989 emit_insn (insns);
3990 valreg = temp;
3993 /* For calls to `setjmp', etc., inform
3994 function.c:setjmp_warnings that it should complain if
3995 nonvolatile values are live. For functions that cannot
3996 return, inform flow that control does not fall through. */
3998 if ((flags & ECF_NORETURN) || pass == 0)
4000 /* The barrier must be emitted
4001 immediately after the CALL_INSN. Some ports emit more
4002 than just a CALL_INSN above, so we must search for it here. */
4004 rtx_insn *last = get_last_insn ();
4005 while (!CALL_P (last))
4007 last = PREV_INSN (last);
4008 /* There was no CALL_INSN? */
4009 gcc_assert (last != before_call);
4012 emit_barrier_after (last);
4014 /* Stack adjustments after a noreturn call are dead code.
4015 However when NO_DEFER_POP is in effect, we must preserve
4016 stack_pointer_delta. */
4017 if (inhibit_defer_pop == 0)
4019 stack_pointer_delta = old_stack_allocated;
4020 pending_stack_adjust = 0;
4024 /* If value type not void, return an rtx for the value. */
4026 if (TYPE_MODE (rettype) == VOIDmode
4027 || ignore)
4028 target = const0_rtx;
4029 else if (structure_value_addr)
4031 if (target == 0 || !MEM_P (target))
4033 target
4034 = gen_rtx_MEM (TYPE_MODE (rettype),
4035 memory_address (TYPE_MODE (rettype),
4036 structure_value_addr));
4037 set_mem_attributes (target, rettype, 1);
4040 else if (pcc_struct_value)
4042 /* This is the special C++ case where we need to
4043 know what the true target was. We take care to
4044 never use this value more than once in one expression. */
4045 target = gen_rtx_MEM (TYPE_MODE (rettype),
4046 copy_to_reg (valreg));
4047 set_mem_attributes (target, rettype, 1);
4049 /* Handle calls that return values in multiple non-contiguous locations.
4050 The Irix 6 ABI has examples of this. */
4051 else if (GET_CODE (valreg) == PARALLEL)
4053 if (target == 0)
4054 target = emit_group_move_into_temps (valreg);
4055 else if (rtx_equal_p (target, valreg))
4057 else if (GET_CODE (target) == PARALLEL)
4058 /* Handle the result of a emit_group_move_into_temps
4059 call in the previous pass. */
4060 emit_group_move (target, valreg);
4061 else
4062 emit_group_store (target, valreg, rettype,
4063 int_size_in_bytes (rettype));
4065 else if (target
4066 && GET_MODE (target) == TYPE_MODE (rettype)
4067 && GET_MODE (target) == GET_MODE (valreg))
4069 bool may_overlap = false;
4071 /* We have to copy a return value in a CLASS_LIKELY_SPILLED hard
4072 reg to a plain register. */
4073 if (!REG_P (target) || HARD_REGISTER_P (target))
4074 valreg = avoid_likely_spilled_reg (valreg);
4076 /* If TARGET is a MEM in the argument area, and we have
4077 saved part of the argument area, then we can't store
4078 directly into TARGET as it may get overwritten when we
4079 restore the argument save area below. Don't work too
4080 hard though and simply force TARGET to a register if it
4081 is a MEM; the optimizer is quite likely to sort it out. */
4082 if (ACCUMULATE_OUTGOING_ARGS && pass && MEM_P (target))
4083 for (i = 0; i < num_actuals; i++)
4084 if (args[i].save_area)
4086 may_overlap = true;
4087 break;
4090 if (may_overlap)
4091 target = copy_to_reg (valreg);
4092 else
4094 /* TARGET and VALREG cannot be equal at this point
4095 because the latter would not have
4096 REG_FUNCTION_VALUE_P true, while the former would if
4097 it were referring to the same register.
4099 If they refer to the same register, this move will be
4100 a no-op, except when function inlining is being
4101 done. */
4102 emit_move_insn (target, valreg);
4104 /* If we are setting a MEM, this code must be executed.
4105 Since it is emitted after the call insn, sibcall
4106 optimization cannot be performed in that case. */
4107 if (MEM_P (target))
4108 sibcall_failure = 1;
4111 else
4112 target = copy_to_reg (avoid_likely_spilled_reg (valreg));
4114 /* If we promoted this return value, make the proper SUBREG.
4115 TARGET might be const0_rtx here, so be careful. */
4116 if (REG_P (target)
4117 && TYPE_MODE (rettype) != BLKmode
4118 && GET_MODE (target) != TYPE_MODE (rettype))
4120 tree type = rettype;
4121 int unsignedp = TYPE_UNSIGNED (type);
4122 int offset = 0;
4123 machine_mode pmode;
4125 /* Ensure we promote as expected, and get the new unsignedness. */
4126 pmode = promote_function_mode (type, TYPE_MODE (type), &unsignedp,
4127 funtype, 1);
4128 gcc_assert (GET_MODE (target) == pmode);
4130 if ((WORDS_BIG_ENDIAN || BYTES_BIG_ENDIAN)
4131 && (GET_MODE_SIZE (GET_MODE (target))
4132 > GET_MODE_SIZE (TYPE_MODE (type))))
4134 offset = GET_MODE_SIZE (GET_MODE (target))
4135 - GET_MODE_SIZE (TYPE_MODE (type));
4136 if (! BYTES_BIG_ENDIAN)
4137 offset = (offset / UNITS_PER_WORD) * UNITS_PER_WORD;
4138 else if (! WORDS_BIG_ENDIAN)
4139 offset %= UNITS_PER_WORD;
4142 target = gen_rtx_SUBREG (TYPE_MODE (type), target, offset);
4143 SUBREG_PROMOTED_VAR_P (target) = 1;
4144 SUBREG_PROMOTED_SET (target, unsignedp);
4147 /* If size of args is variable or this was a constructor call for a stack
4148 argument, restore saved stack-pointer value. */
4150 if (old_stack_level)
4152 rtx_insn *prev = get_last_insn ();
4154 emit_stack_restore (SAVE_BLOCK, old_stack_level);
4155 stack_pointer_delta = old_stack_pointer_delta;
4157 fixup_args_size_notes (prev, get_last_insn (), stack_pointer_delta);
4159 pending_stack_adjust = old_pending_adj;
4160 old_stack_allocated = stack_pointer_delta - pending_stack_adjust;
4161 stack_arg_under_construction = old_stack_arg_under_construction;
4162 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
4163 stack_usage_map = initial_stack_usage_map;
4164 sibcall_failure = 1;
4166 else if (ACCUMULATE_OUTGOING_ARGS && pass)
4168 #ifdef REG_PARM_STACK_SPACE
4169 if (save_area)
4170 restore_fixed_argument_area (save_area, argblock,
4171 high_to_save, low_to_save);
4172 #endif
4174 /* If we saved any argument areas, restore them. */
4175 for (i = 0; i < num_actuals; i++)
4176 if (args[i].save_area)
4178 machine_mode save_mode = GET_MODE (args[i].save_area);
4179 rtx stack_area
4180 = gen_rtx_MEM (save_mode,
4181 memory_address (save_mode,
4182 XEXP (args[i].stack_slot, 0)));
4184 if (save_mode != BLKmode)
4185 emit_move_insn (stack_area, args[i].save_area);
4186 else
4187 emit_block_move (stack_area, args[i].save_area,
4188 GEN_INT (args[i].locate.size.constant),
4189 BLOCK_OP_CALL_PARM);
4192 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
4193 stack_usage_map = initial_stack_usage_map;
4196 /* If this was alloca, record the new stack level. */
4197 if (flags & ECF_MAY_BE_ALLOCA)
4198 record_new_stack_level ();
4200 /* Free up storage we no longer need. */
4201 for (i = 0; i < num_actuals; ++i)
4202 free (args[i].aligned_regs);
4204 targetm.calls.end_call_args ();
4206 insns = get_insns ();
4207 end_sequence ();
4209 if (pass == 0)
4211 tail_call_insns = insns;
4213 /* Restore the pending stack adjustment now that we have
4214 finished generating the sibling call sequence. */
4216 restore_pending_stack_adjust (&save);
4218 /* Prepare arg structure for next iteration. */
4219 for (i = 0; i < num_actuals; i++)
4221 args[i].value = 0;
4222 args[i].aligned_regs = 0;
4223 args[i].stack = 0;
4226 sbitmap_free (stored_args_map);
4227 internal_arg_pointer_exp_state.scan_start = NULL;
4228 internal_arg_pointer_exp_state.cache.release ();
4230 else
4232 normal_call_insns = insns;
4234 /* Verify that we've deallocated all the stack we used. */
4235 gcc_assert ((flags & ECF_NORETURN)
4236 || (old_stack_allocated
4237 == stack_pointer_delta - pending_stack_adjust));
4240 /* If something prevents making this a sibling call,
4241 zero out the sequence. */
4242 if (sibcall_failure)
4243 tail_call_insns = NULL;
4244 else
4245 break;
4248 /* If tail call production succeeded, we need to remove REG_EQUIV notes on
4249 arguments too, as argument area is now clobbered by the call. */
4250 if (tail_call_insns)
4252 emit_insn (tail_call_insns);
4253 crtl->tail_call_emit = true;
4255 else
4257 emit_insn (normal_call_insns);
4258 if (try_tail_call)
4259 /* Ideally we'd emit a message for all of the ways that it could
4260 have failed. */
4261 maybe_complain_about_tail_call (exp, "tail call production failed");
4264 currently_expanding_call--;
4266 free (stack_usage_map_buf);
4267 free (args);
4269 /* Join result with returned bounds so caller may use them if needed. */
4270 target = chkp_join_splitted_slot (target, valbnd);
4272 return target;
4275 /* A sibling call sequence invalidates any REG_EQUIV notes made for
4276 this function's incoming arguments.
4278 At the start of RTL generation we know the only REG_EQUIV notes
4279 in the rtl chain are those for incoming arguments, so we can look
4280 for REG_EQUIV notes between the start of the function and the
4281 NOTE_INSN_FUNCTION_BEG.
4283 This is (slight) overkill. We could keep track of the highest
4284 argument we clobber and be more selective in removing notes, but it
4285 does not seem to be worth the effort. */
4287 void
4288 fixup_tail_calls (void)
4290 rtx_insn *insn;
4292 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
4294 rtx note;
4296 /* There are never REG_EQUIV notes for the incoming arguments
4297 after the NOTE_INSN_FUNCTION_BEG note, so stop if we see it. */
4298 if (NOTE_P (insn)
4299 && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG)
4300 break;
4302 note = find_reg_note (insn, REG_EQUIV, 0);
4303 if (note)
4304 remove_note (insn, note);
4305 note = find_reg_note (insn, REG_EQUIV, 0);
4306 gcc_assert (!note);
4310 /* Traverse a list of TYPES and expand all complex types into their
4311 components. */
4312 static tree
4313 split_complex_types (tree types)
4315 tree p;
4317 /* Before allocating memory, check for the common case of no complex. */
4318 for (p = types; p; p = TREE_CHAIN (p))
4320 tree type = TREE_VALUE (p);
4321 if (TREE_CODE (type) == COMPLEX_TYPE
4322 && targetm.calls.split_complex_arg (type))
4323 goto found;
4325 return types;
4327 found:
4328 types = copy_list (types);
4330 for (p = types; p; p = TREE_CHAIN (p))
4332 tree complex_type = TREE_VALUE (p);
4334 if (TREE_CODE (complex_type) == COMPLEX_TYPE
4335 && targetm.calls.split_complex_arg (complex_type))
4337 tree next, imag;
4339 /* Rewrite complex type with component type. */
4340 TREE_VALUE (p) = TREE_TYPE (complex_type);
4341 next = TREE_CHAIN (p);
4343 /* Add another component type for the imaginary part. */
4344 imag = build_tree_list (NULL_TREE, TREE_VALUE (p));
4345 TREE_CHAIN (p) = imag;
4346 TREE_CHAIN (imag) = next;
4348 /* Skip the newly created node. */
4349 p = TREE_CHAIN (p);
4353 return types;
4356 /* Output a library call to function FUN (a SYMBOL_REF rtx).
4357 The RETVAL parameter specifies whether return value needs to be saved, other
4358 parameters are documented in the emit_library_call function below. */
4360 static rtx
4361 emit_library_call_value_1 (int retval, rtx orgfun, rtx value,
4362 enum libcall_type fn_type,
4363 machine_mode outmode, int nargs, va_list p)
4365 /* Total size in bytes of all the stack-parms scanned so far. */
4366 struct args_size args_size;
4367 /* Size of arguments before any adjustments (such as rounding). */
4368 struct args_size original_args_size;
4369 int argnum;
4370 rtx fun;
4371 /* Todo, choose the correct decl type of orgfun. Sadly this information
4372 isn't present here, so we default to native calling abi here. */
4373 tree fndecl ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */
4374 tree fntype ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */
4375 int count;
4376 rtx argblock = 0;
4377 CUMULATIVE_ARGS args_so_far_v;
4378 cumulative_args_t args_so_far;
4379 struct arg
4381 rtx value;
4382 machine_mode mode;
4383 rtx reg;
4384 int partial;
4385 struct locate_and_pad_arg_data locate;
4386 rtx save_area;
4388 struct arg *argvec;
4389 int old_inhibit_defer_pop = inhibit_defer_pop;
4390 rtx call_fusage = 0;
4391 rtx mem_value = 0;
4392 rtx valreg;
4393 int pcc_struct_value = 0;
4394 int struct_value_size = 0;
4395 int flags;
4396 int reg_parm_stack_space = 0;
4397 int needed;
4398 rtx_insn *before_call;
4399 bool have_push_fusage;
4400 tree tfom; /* type_for_mode (outmode, 0) */
4402 #ifdef REG_PARM_STACK_SPACE
4403 /* Define the boundary of the register parm stack space that needs to be
4404 save, if any. */
4405 int low_to_save = 0, high_to_save = 0;
4406 rtx save_area = 0; /* Place that it is saved. */
4407 #endif
4409 /* Size of the stack reserved for parameter registers. */
4410 int initial_highest_arg_in_use = highest_outgoing_arg_in_use;
4411 char *initial_stack_usage_map = stack_usage_map;
4412 char *stack_usage_map_buf = NULL;
4414 rtx struct_value = targetm.calls.struct_value_rtx (0, 0);
4416 #ifdef REG_PARM_STACK_SPACE
4417 reg_parm_stack_space = REG_PARM_STACK_SPACE ((tree) 0);
4418 #endif
4420 /* By default, library functions cannot throw. */
4421 flags = ECF_NOTHROW;
4423 switch (fn_type)
4425 case LCT_NORMAL:
4426 break;
4427 case LCT_CONST:
4428 flags |= ECF_CONST;
4429 break;
4430 case LCT_PURE:
4431 flags |= ECF_PURE;
4432 break;
4433 case LCT_NORETURN:
4434 flags |= ECF_NORETURN;
4435 break;
4436 case LCT_THROW:
4437 flags &= ~ECF_NOTHROW;
4438 break;
4439 case LCT_RETURNS_TWICE:
4440 flags = ECF_RETURNS_TWICE;
4441 break;
4443 fun = orgfun;
4445 /* Ensure current function's preferred stack boundary is at least
4446 what we need. */
4447 if (crtl->preferred_stack_boundary < PREFERRED_STACK_BOUNDARY)
4448 crtl->preferred_stack_boundary = PREFERRED_STACK_BOUNDARY;
4450 /* If this kind of value comes back in memory,
4451 decide where in memory it should come back. */
4452 if (outmode != VOIDmode)
4454 tfom = lang_hooks.types.type_for_mode (outmode, 0);
4455 if (aggregate_value_p (tfom, 0))
4457 #ifdef PCC_STATIC_STRUCT_RETURN
4458 rtx pointer_reg
4459 = hard_function_value (build_pointer_type (tfom), 0, 0, 0);
4460 mem_value = gen_rtx_MEM (outmode, pointer_reg);
4461 pcc_struct_value = 1;
4462 if (value == 0)
4463 value = gen_reg_rtx (outmode);
4464 #else /* not PCC_STATIC_STRUCT_RETURN */
4465 struct_value_size = GET_MODE_SIZE (outmode);
4466 if (value != 0 && MEM_P (value))
4467 mem_value = value;
4468 else
4469 mem_value = assign_temp (tfom, 1, 1);
4470 #endif
4471 /* This call returns a big structure. */
4472 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
4475 else
4476 tfom = void_type_node;
4478 /* ??? Unfinished: must pass the memory address as an argument. */
4480 /* Copy all the libcall-arguments out of the varargs data
4481 and into a vector ARGVEC.
4483 Compute how to pass each argument. We only support a very small subset
4484 of the full argument passing conventions to limit complexity here since
4485 library functions shouldn't have many args. */
4487 argvec = XALLOCAVEC (struct arg, nargs + 1);
4488 memset (argvec, 0, (nargs + 1) * sizeof (struct arg));
4490 #ifdef INIT_CUMULATIVE_LIBCALL_ARGS
4491 INIT_CUMULATIVE_LIBCALL_ARGS (args_so_far_v, outmode, fun);
4492 #else
4493 INIT_CUMULATIVE_ARGS (args_so_far_v, NULL_TREE, fun, 0, nargs);
4494 #endif
4495 args_so_far = pack_cumulative_args (&args_so_far_v);
4497 args_size.constant = 0;
4498 args_size.var = 0;
4500 count = 0;
4502 push_temp_slots ();
4504 /* If there's a structure value address to be passed,
4505 either pass it in the special place, or pass it as an extra argument. */
4506 if (mem_value && struct_value == 0 && ! pcc_struct_value)
4508 rtx addr = XEXP (mem_value, 0);
4510 nargs++;
4512 /* Make sure it is a reasonable operand for a move or push insn. */
4513 if (!REG_P (addr) && !MEM_P (addr)
4514 && !(CONSTANT_P (addr)
4515 && targetm.legitimate_constant_p (Pmode, addr)))
4516 addr = force_operand (addr, NULL_RTX);
4518 argvec[count].value = addr;
4519 argvec[count].mode = Pmode;
4520 argvec[count].partial = 0;
4522 argvec[count].reg = targetm.calls.function_arg (args_so_far,
4523 Pmode, NULL_TREE, true);
4524 gcc_assert (targetm.calls.arg_partial_bytes (args_so_far, Pmode,
4525 NULL_TREE, 1) == 0);
4527 locate_and_pad_parm (Pmode, NULL_TREE,
4528 #ifdef STACK_PARMS_IN_REG_PARM_AREA
4530 #else
4531 argvec[count].reg != 0,
4532 #endif
4533 reg_parm_stack_space, 0,
4534 NULL_TREE, &args_size, &argvec[count].locate);
4536 if (argvec[count].reg == 0 || argvec[count].partial != 0
4537 || reg_parm_stack_space > 0)
4538 args_size.constant += argvec[count].locate.size.constant;
4540 targetm.calls.function_arg_advance (args_so_far, Pmode, (tree) 0, true);
4542 count++;
4545 for (; count < nargs; count++)
4547 rtx val = va_arg (p, rtx);
4548 machine_mode mode = (machine_mode) va_arg (p, int);
4549 int unsigned_p = 0;
4551 /* We cannot convert the arg value to the mode the library wants here;
4552 must do it earlier where we know the signedness of the arg. */
4553 gcc_assert (mode != BLKmode
4554 && (GET_MODE (val) == mode || GET_MODE (val) == VOIDmode));
4556 /* Make sure it is a reasonable operand for a move or push insn. */
4557 if (!REG_P (val) && !MEM_P (val)
4558 && !(CONSTANT_P (val) && targetm.legitimate_constant_p (mode, val)))
4559 val = force_operand (val, NULL_RTX);
4561 if (pass_by_reference (&args_so_far_v, mode, NULL_TREE, 1))
4563 rtx slot;
4564 int must_copy
4565 = !reference_callee_copied (&args_so_far_v, mode, NULL_TREE, 1);
4567 /* If this was a CONST function, it is now PURE since it now
4568 reads memory. */
4569 if (flags & ECF_CONST)
4571 flags &= ~ECF_CONST;
4572 flags |= ECF_PURE;
4575 if (MEM_P (val) && !must_copy)
4577 tree val_expr = MEM_EXPR (val);
4578 if (val_expr)
4579 mark_addressable (val_expr);
4580 slot = val;
4582 else
4584 slot = assign_temp (lang_hooks.types.type_for_mode (mode, 0),
4585 1, 1);
4586 emit_move_insn (slot, val);
4589 call_fusage = gen_rtx_EXPR_LIST (VOIDmode,
4590 gen_rtx_USE (VOIDmode, slot),
4591 call_fusage);
4592 if (must_copy)
4593 call_fusage = gen_rtx_EXPR_LIST (VOIDmode,
4594 gen_rtx_CLOBBER (VOIDmode,
4595 slot),
4596 call_fusage);
4598 mode = Pmode;
4599 val = force_operand (XEXP (slot, 0), NULL_RTX);
4602 mode = promote_function_mode (NULL_TREE, mode, &unsigned_p, NULL_TREE, 0);
4603 argvec[count].mode = mode;
4604 argvec[count].value = convert_modes (mode, GET_MODE (val), val, unsigned_p);
4605 argvec[count].reg = targetm.calls.function_arg (args_so_far, mode,
4606 NULL_TREE, true);
4608 argvec[count].partial
4609 = targetm.calls.arg_partial_bytes (args_so_far, mode, NULL_TREE, 1);
4611 if (argvec[count].reg == 0
4612 || argvec[count].partial != 0
4613 || reg_parm_stack_space > 0)
4615 locate_and_pad_parm (mode, NULL_TREE,
4616 #ifdef STACK_PARMS_IN_REG_PARM_AREA
4618 #else
4619 argvec[count].reg != 0,
4620 #endif
4621 reg_parm_stack_space, argvec[count].partial,
4622 NULL_TREE, &args_size, &argvec[count].locate);
4623 args_size.constant += argvec[count].locate.size.constant;
4624 gcc_assert (!argvec[count].locate.size.var);
4626 #ifdef BLOCK_REG_PADDING
4627 else
4628 /* The argument is passed entirely in registers. See at which
4629 end it should be padded. */
4630 argvec[count].locate.where_pad =
4631 BLOCK_REG_PADDING (mode, NULL_TREE,
4632 GET_MODE_SIZE (mode) <= UNITS_PER_WORD);
4633 #endif
4635 targetm.calls.function_arg_advance (args_so_far, mode, (tree) 0, true);
4638 /* If this machine requires an external definition for library
4639 functions, write one out. */
4640 assemble_external_libcall (fun);
4642 original_args_size = args_size;
4643 args_size.constant = (((args_size.constant
4644 + stack_pointer_delta
4645 + STACK_BYTES - 1)
4646 / STACK_BYTES
4647 * STACK_BYTES)
4648 - stack_pointer_delta);
4650 args_size.constant = MAX (args_size.constant,
4651 reg_parm_stack_space);
4653 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
4654 args_size.constant -= reg_parm_stack_space;
4656 if (args_size.constant > crtl->outgoing_args_size)
4657 crtl->outgoing_args_size = args_size.constant;
4659 if (flag_stack_usage_info && !ACCUMULATE_OUTGOING_ARGS)
4661 int pushed = args_size.constant + pending_stack_adjust;
4662 if (pushed > current_function_pushed_stack_size)
4663 current_function_pushed_stack_size = pushed;
4666 if (ACCUMULATE_OUTGOING_ARGS)
4668 /* Since the stack pointer will never be pushed, it is possible for
4669 the evaluation of a parm to clobber something we have already
4670 written to the stack. Since most function calls on RISC machines
4671 do not use the stack, this is uncommon, but must work correctly.
4673 Therefore, we save any area of the stack that was already written
4674 and that we are using. Here we set up to do this by making a new
4675 stack usage map from the old one.
4677 Another approach might be to try to reorder the argument
4678 evaluations to avoid this conflicting stack usage. */
4680 needed = args_size.constant;
4682 /* Since we will be writing into the entire argument area, the
4683 map must be allocated for its entire size, not just the part that
4684 is the responsibility of the caller. */
4685 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
4686 needed += reg_parm_stack_space;
4688 if (ARGS_GROW_DOWNWARD)
4689 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use,
4690 needed + 1);
4691 else
4692 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use, needed);
4694 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use);
4695 stack_usage_map = stack_usage_map_buf;
4697 if (initial_highest_arg_in_use)
4698 memcpy (stack_usage_map, initial_stack_usage_map,
4699 initial_highest_arg_in_use);
4701 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use)
4702 memset (&stack_usage_map[initial_highest_arg_in_use], 0,
4703 highest_outgoing_arg_in_use - initial_highest_arg_in_use);
4704 needed = 0;
4706 /* We must be careful to use virtual regs before they're instantiated,
4707 and real regs afterwards. Loop optimization, for example, can create
4708 new libcalls after we've instantiated the virtual regs, and if we
4709 use virtuals anyway, they won't match the rtl patterns. */
4711 if (virtuals_instantiated)
4712 argblock = plus_constant (Pmode, stack_pointer_rtx,
4713 STACK_POINTER_OFFSET);
4714 else
4715 argblock = virtual_outgoing_args_rtx;
4717 else
4719 if (!PUSH_ARGS)
4720 argblock = push_block (GEN_INT (args_size.constant), 0, 0);
4723 /* We push args individually in reverse order, perform stack alignment
4724 before the first push (the last arg). */
4725 if (argblock == 0)
4726 anti_adjust_stack (GEN_INT (args_size.constant
4727 - original_args_size.constant));
4729 argnum = nargs - 1;
4731 #ifdef REG_PARM_STACK_SPACE
4732 if (ACCUMULATE_OUTGOING_ARGS)
4734 /* The argument list is the property of the called routine and it
4735 may clobber it. If the fixed area has been used for previous
4736 parameters, we must save and restore it. */
4737 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock,
4738 &low_to_save, &high_to_save);
4740 #endif
4742 /* When expanding a normal call, args are stored in push order,
4743 which is the reverse of what we have here. */
4744 bool any_regs = false;
4745 for (int i = nargs; i-- > 0; )
4746 if (argvec[i].reg != NULL_RTX)
4748 targetm.calls.call_args (argvec[i].reg, NULL_TREE);
4749 any_regs = true;
4751 if (!any_regs)
4752 targetm.calls.call_args (pc_rtx, NULL_TREE);
4754 /* Push the args that need to be pushed. */
4756 have_push_fusage = false;
4758 /* ARGNUM indexes the ARGVEC array in the order in which the arguments
4759 are to be pushed. */
4760 for (count = 0; count < nargs; count++, argnum--)
4762 machine_mode mode = argvec[argnum].mode;
4763 rtx val = argvec[argnum].value;
4764 rtx reg = argvec[argnum].reg;
4765 int partial = argvec[argnum].partial;
4766 unsigned int parm_align = argvec[argnum].locate.boundary;
4767 int lower_bound = 0, upper_bound = 0, i;
4769 if (! (reg != 0 && partial == 0))
4771 rtx use;
4773 if (ACCUMULATE_OUTGOING_ARGS)
4775 /* If this is being stored into a pre-allocated, fixed-size,
4776 stack area, save any previous data at that location. */
4778 if (ARGS_GROW_DOWNWARD)
4780 /* stack_slot is negative, but we want to index stack_usage_map
4781 with positive values. */
4782 upper_bound = -argvec[argnum].locate.slot_offset.constant + 1;
4783 lower_bound = upper_bound - argvec[argnum].locate.size.constant;
4785 else
4787 lower_bound = argvec[argnum].locate.slot_offset.constant;
4788 upper_bound = lower_bound + argvec[argnum].locate.size.constant;
4791 i = lower_bound;
4792 /* Don't worry about things in the fixed argument area;
4793 it has already been saved. */
4794 if (i < reg_parm_stack_space)
4795 i = reg_parm_stack_space;
4796 while (i < upper_bound && stack_usage_map[i] == 0)
4797 i++;
4799 if (i < upper_bound)
4801 /* We need to make a save area. */
4802 unsigned int size
4803 = argvec[argnum].locate.size.constant * BITS_PER_UNIT;
4804 machine_mode save_mode
4805 = mode_for_size (size, MODE_INT, 1);
4806 rtx adr
4807 = plus_constant (Pmode, argblock,
4808 argvec[argnum].locate.offset.constant);
4809 rtx stack_area
4810 = gen_rtx_MEM (save_mode, memory_address (save_mode, adr));
4812 if (save_mode == BLKmode)
4814 argvec[argnum].save_area
4815 = assign_stack_temp (BLKmode,
4816 argvec[argnum].locate.size.constant
4819 emit_block_move (validize_mem
4820 (copy_rtx (argvec[argnum].save_area)),
4821 stack_area,
4822 GEN_INT (argvec[argnum].locate.size.constant),
4823 BLOCK_OP_CALL_PARM);
4825 else
4827 argvec[argnum].save_area = gen_reg_rtx (save_mode);
4829 emit_move_insn (argvec[argnum].save_area, stack_area);
4834 emit_push_insn (val, mode, NULL_TREE, NULL_RTX, parm_align,
4835 partial, reg, 0, argblock,
4836 GEN_INT (argvec[argnum].locate.offset.constant),
4837 reg_parm_stack_space,
4838 ARGS_SIZE_RTX (argvec[argnum].locate.alignment_pad), false);
4840 /* Now mark the segment we just used. */
4841 if (ACCUMULATE_OUTGOING_ARGS)
4842 for (i = lower_bound; i < upper_bound; i++)
4843 stack_usage_map[i] = 1;
4845 NO_DEFER_POP;
4847 /* Indicate argument access so that alias.c knows that these
4848 values are live. */
4849 if (argblock)
4850 use = plus_constant (Pmode, argblock,
4851 argvec[argnum].locate.offset.constant);
4852 else if (have_push_fusage)
4853 continue;
4854 else
4856 /* When arguments are pushed, trying to tell alias.c where
4857 exactly this argument is won't work, because the
4858 auto-increment causes confusion. So we merely indicate
4859 that we access something with a known mode somewhere on
4860 the stack. */
4861 use = gen_rtx_PLUS (Pmode, stack_pointer_rtx,
4862 gen_rtx_SCRATCH (Pmode));
4863 have_push_fusage = true;
4865 use = gen_rtx_MEM (argvec[argnum].mode, use);
4866 use = gen_rtx_USE (VOIDmode, use);
4867 call_fusage = gen_rtx_EXPR_LIST (VOIDmode, use, call_fusage);
4871 argnum = nargs - 1;
4873 fun = prepare_call_address (NULL, fun, NULL, &call_fusage, 0, 0);
4875 /* Now load any reg parms into their regs. */
4877 /* ARGNUM indexes the ARGVEC array in the order in which the arguments
4878 are to be pushed. */
4879 for (count = 0; count < nargs; count++, argnum--)
4881 machine_mode mode = argvec[argnum].mode;
4882 rtx val = argvec[argnum].value;
4883 rtx reg = argvec[argnum].reg;
4884 int partial = argvec[argnum].partial;
4885 #ifdef BLOCK_REG_PADDING
4886 int size = 0;
4887 #endif
4889 /* Handle calls that pass values in multiple non-contiguous
4890 locations. The PA64 has examples of this for library calls. */
4891 if (reg != 0 && GET_CODE (reg) == PARALLEL)
4892 emit_group_load (reg, val, NULL_TREE, GET_MODE_SIZE (mode));
4893 else if (reg != 0 && partial == 0)
4895 emit_move_insn (reg, val);
4896 #ifdef BLOCK_REG_PADDING
4897 size = GET_MODE_SIZE (argvec[argnum].mode);
4899 /* Copied from load_register_parameters. */
4901 /* Handle case where we have a value that needs shifting
4902 up to the msb. eg. a QImode value and we're padding
4903 upward on a BYTES_BIG_ENDIAN machine. */
4904 if (size < UNITS_PER_WORD
4905 && (argvec[argnum].locate.where_pad
4906 == (BYTES_BIG_ENDIAN ? upward : downward)))
4908 rtx x;
4909 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
4911 /* Assigning REG here rather than a temp makes CALL_FUSAGE
4912 report the whole reg as used. Strictly speaking, the
4913 call only uses SIZE bytes at the msb end, but it doesn't
4914 seem worth generating rtl to say that. */
4915 reg = gen_rtx_REG (word_mode, REGNO (reg));
4916 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1);
4917 if (x != reg)
4918 emit_move_insn (reg, x);
4920 #endif
4923 NO_DEFER_POP;
4926 /* Any regs containing parms remain in use through the call. */
4927 for (count = 0; count < nargs; count++)
4929 rtx reg = argvec[count].reg;
4930 if (reg != 0 && GET_CODE (reg) == PARALLEL)
4931 use_group_regs (&call_fusage, reg);
4932 else if (reg != 0)
4934 int partial = argvec[count].partial;
4935 if (partial)
4937 int nregs;
4938 gcc_assert (partial % UNITS_PER_WORD == 0);
4939 nregs = partial / UNITS_PER_WORD;
4940 use_regs (&call_fusage, REGNO (reg), nregs);
4942 else
4943 use_reg (&call_fusage, reg);
4947 /* Pass the function the address in which to return a structure value. */
4948 if (mem_value != 0 && struct_value != 0 && ! pcc_struct_value)
4950 emit_move_insn (struct_value,
4951 force_reg (Pmode,
4952 force_operand (XEXP (mem_value, 0),
4953 NULL_RTX)));
4954 if (REG_P (struct_value))
4955 use_reg (&call_fusage, struct_value);
4958 /* Don't allow popping to be deferred, since then
4959 cse'ing of library calls could delete a call and leave the pop. */
4960 NO_DEFER_POP;
4961 valreg = (mem_value == 0 && outmode != VOIDmode
4962 ? hard_libcall_value (outmode, orgfun) : NULL_RTX);
4964 /* Stack must be properly aligned now. */
4965 gcc_assert (!(stack_pointer_delta
4966 & (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT - 1)));
4968 before_call = get_last_insn ();
4970 /* We pass the old value of inhibit_defer_pop + 1 to emit_call_1, which
4971 will set inhibit_defer_pop to that value. */
4972 /* The return type is needed to decide how many bytes the function pops.
4973 Signedness plays no role in that, so for simplicity, we pretend it's
4974 always signed. We also assume that the list of arguments passed has
4975 no impact, so we pretend it is unknown. */
4977 emit_call_1 (fun, NULL,
4978 get_identifier (XSTR (orgfun, 0)),
4979 build_function_type (tfom, NULL_TREE),
4980 original_args_size.constant, args_size.constant,
4981 struct_value_size,
4982 targetm.calls.function_arg (args_so_far,
4983 VOIDmode, void_type_node, true),
4984 valreg,
4985 old_inhibit_defer_pop + 1, call_fusage, flags, args_so_far);
4987 if (flag_ipa_ra)
4989 rtx datum = orgfun;
4990 gcc_assert (GET_CODE (datum) == SYMBOL_REF);
4991 rtx_call_insn *last = last_call_insn ();
4992 add_reg_note (last, REG_CALL_DECL, datum);
4995 /* Right-shift returned value if necessary. */
4996 if (!pcc_struct_value
4997 && TYPE_MODE (tfom) != BLKmode
4998 && targetm.calls.return_in_msb (tfom))
5000 shift_return_value (TYPE_MODE (tfom), false, valreg);
5001 valreg = gen_rtx_REG (TYPE_MODE (tfom), REGNO (valreg));
5004 targetm.calls.end_call_args ();
5006 /* For calls to `setjmp', etc., inform function.c:setjmp_warnings
5007 that it should complain if nonvolatile values are live. For
5008 functions that cannot return, inform flow that control does not
5009 fall through. */
5010 if (flags & ECF_NORETURN)
5012 /* The barrier note must be emitted
5013 immediately after the CALL_INSN. Some ports emit more than
5014 just a CALL_INSN above, so we must search for it here. */
5015 rtx_insn *last = get_last_insn ();
5016 while (!CALL_P (last))
5018 last = PREV_INSN (last);
5019 /* There was no CALL_INSN? */
5020 gcc_assert (last != before_call);
5023 emit_barrier_after (last);
5026 /* Consider that "regular" libcalls, i.e. all of them except for LCT_THROW
5027 and LCT_RETURNS_TWICE, cannot perform non-local gotos. */
5028 if (flags & ECF_NOTHROW)
5030 rtx_insn *last = get_last_insn ();
5031 while (!CALL_P (last))
5033 last = PREV_INSN (last);
5034 /* There was no CALL_INSN? */
5035 gcc_assert (last != before_call);
5038 make_reg_eh_region_note_nothrow_nononlocal (last);
5041 /* Now restore inhibit_defer_pop to its actual original value. */
5042 OK_DEFER_POP;
5044 pop_temp_slots ();
5046 /* Copy the value to the right place. */
5047 if (outmode != VOIDmode && retval)
5049 if (mem_value)
5051 if (value == 0)
5052 value = mem_value;
5053 if (value != mem_value)
5054 emit_move_insn (value, mem_value);
5056 else if (GET_CODE (valreg) == PARALLEL)
5058 if (value == 0)
5059 value = gen_reg_rtx (outmode);
5060 emit_group_store (value, valreg, NULL_TREE, GET_MODE_SIZE (outmode));
5062 else
5064 /* Convert to the proper mode if a promotion has been active. */
5065 if (GET_MODE (valreg) != outmode)
5067 int unsignedp = TYPE_UNSIGNED (tfom);
5069 gcc_assert (promote_function_mode (tfom, outmode, &unsignedp,
5070 fndecl ? TREE_TYPE (fndecl) : fntype, 1)
5071 == GET_MODE (valreg));
5072 valreg = convert_modes (outmode, GET_MODE (valreg), valreg, 0);
5075 if (value != 0)
5076 emit_move_insn (value, valreg);
5077 else
5078 value = valreg;
5082 if (ACCUMULATE_OUTGOING_ARGS)
5084 #ifdef REG_PARM_STACK_SPACE
5085 if (save_area)
5086 restore_fixed_argument_area (save_area, argblock,
5087 high_to_save, low_to_save);
5088 #endif
5090 /* If we saved any argument areas, restore them. */
5091 for (count = 0; count < nargs; count++)
5092 if (argvec[count].save_area)
5094 machine_mode save_mode = GET_MODE (argvec[count].save_area);
5095 rtx adr = plus_constant (Pmode, argblock,
5096 argvec[count].locate.offset.constant);
5097 rtx stack_area = gen_rtx_MEM (save_mode,
5098 memory_address (save_mode, adr));
5100 if (save_mode == BLKmode)
5101 emit_block_move (stack_area,
5102 validize_mem
5103 (copy_rtx (argvec[count].save_area)),
5104 GEN_INT (argvec[count].locate.size.constant),
5105 BLOCK_OP_CALL_PARM);
5106 else
5107 emit_move_insn (stack_area, argvec[count].save_area);
5110 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
5111 stack_usage_map = initial_stack_usage_map;
5114 free (stack_usage_map_buf);
5116 return value;
5120 /* Output a library call to function FUN (a SYMBOL_REF rtx)
5121 (emitting the queue unless NO_QUEUE is nonzero),
5122 for a value of mode OUTMODE,
5123 with NARGS different arguments, passed as alternating rtx values
5124 and machine_modes to convert them to.
5126 FN_TYPE should be LCT_NORMAL for `normal' calls, LCT_CONST for
5127 `const' calls, LCT_PURE for `pure' calls, or other LCT_ value for
5128 other types of library calls. */
5130 void
5131 emit_library_call (rtx orgfun, enum libcall_type fn_type,
5132 machine_mode outmode, int nargs, ...)
5134 va_list p;
5136 va_start (p, nargs);
5137 emit_library_call_value_1 (0, orgfun, NULL_RTX, fn_type, outmode, nargs, p);
5138 va_end (p);
5141 /* Like emit_library_call except that an extra argument, VALUE,
5142 comes second and says where to store the result.
5143 (If VALUE is zero, this function chooses a convenient way
5144 to return the value.
5146 This function returns an rtx for where the value is to be found.
5147 If VALUE is nonzero, VALUE is returned. */
5150 emit_library_call_value (rtx orgfun, rtx value,
5151 enum libcall_type fn_type,
5152 machine_mode outmode, int nargs, ...)
5154 rtx result;
5155 va_list p;
5157 va_start (p, nargs);
5158 result = emit_library_call_value_1 (1, orgfun, value, fn_type, outmode,
5159 nargs, p);
5160 va_end (p);
5162 return result;
5166 /* Store pointer bounds argument ARG into Bounds Table entry
5167 associated with PARM. */
5168 static void
5169 store_bounds (struct arg_data *arg, struct arg_data *parm)
5171 rtx slot = NULL, ptr = NULL, addr = NULL;
5173 /* We may pass bounds not associated with any pointer. */
5174 if (!parm)
5176 gcc_assert (arg->special_slot);
5177 slot = arg->special_slot;
5178 ptr = const0_rtx;
5180 /* Find pointer associated with bounds and where it is
5181 passed. */
5182 else
5184 if (!parm->reg)
5186 gcc_assert (!arg->special_slot);
5188 addr = adjust_address (parm->stack, Pmode, arg->pointer_offset);
5190 else if (REG_P (parm->reg))
5192 gcc_assert (arg->special_slot);
5193 slot = arg->special_slot;
5195 if (MEM_P (parm->value))
5196 addr = adjust_address (parm->value, Pmode, arg->pointer_offset);
5197 else if (REG_P (parm->value))
5198 ptr = gen_rtx_SUBREG (Pmode, parm->value, arg->pointer_offset);
5199 else
5201 gcc_assert (!arg->pointer_offset);
5202 ptr = parm->value;
5205 else
5207 gcc_assert (GET_CODE (parm->reg) == PARALLEL);
5209 gcc_assert (arg->special_slot);
5210 slot = arg->special_slot;
5212 if (parm->parallel_value)
5213 ptr = chkp_get_value_with_offs (parm->parallel_value,
5214 GEN_INT (arg->pointer_offset));
5215 else
5216 gcc_unreachable ();
5220 /* Expand bounds. */
5221 if (!arg->value)
5222 arg->value = expand_normal (arg->tree_value);
5224 targetm.calls.store_bounds_for_arg (ptr, addr, arg->value, slot);
5227 /* Store a single argument for a function call
5228 into the register or memory area where it must be passed.
5229 *ARG describes the argument value and where to pass it.
5231 ARGBLOCK is the address of the stack-block for all the arguments,
5232 or 0 on a machine where arguments are pushed individually.
5234 MAY_BE_ALLOCA nonzero says this could be a call to `alloca'
5235 so must be careful about how the stack is used.
5237 VARIABLE_SIZE nonzero says that this was a variable-sized outgoing
5238 argument stack. This is used if ACCUMULATE_OUTGOING_ARGS to indicate
5239 that we need not worry about saving and restoring the stack.
5241 FNDECL is the declaration of the function we are calling.
5243 Return nonzero if this arg should cause sibcall failure,
5244 zero otherwise. */
5246 static int
5247 store_one_arg (struct arg_data *arg, rtx argblock, int flags,
5248 int variable_size ATTRIBUTE_UNUSED, int reg_parm_stack_space)
5250 tree pval = arg->tree_value;
5251 rtx reg = 0;
5252 int partial = 0;
5253 int used = 0;
5254 int i, lower_bound = 0, upper_bound = 0;
5255 int sibcall_failure = 0;
5257 if (TREE_CODE (pval) == ERROR_MARK)
5258 return 1;
5260 /* Push a new temporary level for any temporaries we make for
5261 this argument. */
5262 push_temp_slots ();
5264 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL))
5266 /* If this is being stored into a pre-allocated, fixed-size, stack area,
5267 save any previous data at that location. */
5268 if (argblock && ! variable_size && arg->stack)
5270 if (ARGS_GROW_DOWNWARD)
5272 /* stack_slot is negative, but we want to index stack_usage_map
5273 with positive values. */
5274 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS)
5275 upper_bound = -INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1)) + 1;
5276 else
5277 upper_bound = 0;
5279 lower_bound = upper_bound - arg->locate.size.constant;
5281 else
5283 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS)
5284 lower_bound = INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1));
5285 else
5286 lower_bound = 0;
5288 upper_bound = lower_bound + arg->locate.size.constant;
5291 i = lower_bound;
5292 /* Don't worry about things in the fixed argument area;
5293 it has already been saved. */
5294 if (i < reg_parm_stack_space)
5295 i = reg_parm_stack_space;
5296 while (i < upper_bound && stack_usage_map[i] == 0)
5297 i++;
5299 if (i < upper_bound)
5301 /* We need to make a save area. */
5302 unsigned int size = arg->locate.size.constant * BITS_PER_UNIT;
5303 machine_mode save_mode = mode_for_size (size, MODE_INT, 1);
5304 rtx adr = memory_address (save_mode, XEXP (arg->stack_slot, 0));
5305 rtx stack_area = gen_rtx_MEM (save_mode, adr);
5307 if (save_mode == BLKmode)
5309 arg->save_area
5310 = assign_temp (TREE_TYPE (arg->tree_value), 1, 1);
5311 preserve_temp_slots (arg->save_area);
5312 emit_block_move (validize_mem (copy_rtx (arg->save_area)),
5313 stack_area,
5314 GEN_INT (arg->locate.size.constant),
5315 BLOCK_OP_CALL_PARM);
5317 else
5319 arg->save_area = gen_reg_rtx (save_mode);
5320 emit_move_insn (arg->save_area, stack_area);
5326 /* If this isn't going to be placed on both the stack and in registers,
5327 set up the register and number of words. */
5328 if (! arg->pass_on_stack)
5330 if (flags & ECF_SIBCALL)
5331 reg = arg->tail_call_reg;
5332 else
5333 reg = arg->reg;
5334 partial = arg->partial;
5337 /* Being passed entirely in a register. We shouldn't be called in
5338 this case. */
5339 gcc_assert (reg == 0 || partial != 0);
5341 /* If this arg needs special alignment, don't load the registers
5342 here. */
5343 if (arg->n_aligned_regs != 0)
5344 reg = 0;
5346 /* If this is being passed partially in a register, we can't evaluate
5347 it directly into its stack slot. Otherwise, we can. */
5348 if (arg->value == 0)
5350 /* stack_arg_under_construction is nonzero if a function argument is
5351 being evaluated directly into the outgoing argument list and
5352 expand_call must take special action to preserve the argument list
5353 if it is called recursively.
5355 For scalar function arguments stack_usage_map is sufficient to
5356 determine which stack slots must be saved and restored. Scalar
5357 arguments in general have pass_on_stack == 0.
5359 If this argument is initialized by a function which takes the
5360 address of the argument (a C++ constructor or a C function
5361 returning a BLKmode structure), then stack_usage_map is
5362 insufficient and expand_call must push the stack around the
5363 function call. Such arguments have pass_on_stack == 1.
5365 Note that it is always safe to set stack_arg_under_construction,
5366 but this generates suboptimal code if set when not needed. */
5368 if (arg->pass_on_stack)
5369 stack_arg_under_construction++;
5371 arg->value = expand_expr (pval,
5372 (partial
5373 || TYPE_MODE (TREE_TYPE (pval)) != arg->mode)
5374 ? NULL_RTX : arg->stack,
5375 VOIDmode, EXPAND_STACK_PARM);
5377 /* If we are promoting object (or for any other reason) the mode
5378 doesn't agree, convert the mode. */
5380 if (arg->mode != TYPE_MODE (TREE_TYPE (pval)))
5381 arg->value = convert_modes (arg->mode, TYPE_MODE (TREE_TYPE (pval)),
5382 arg->value, arg->unsignedp);
5384 if (arg->pass_on_stack)
5385 stack_arg_under_construction--;
5388 /* Check for overlap with already clobbered argument area. */
5389 if ((flags & ECF_SIBCALL)
5390 && MEM_P (arg->value)
5391 && mem_overlaps_already_clobbered_arg_p (XEXP (arg->value, 0),
5392 arg->locate.size.constant))
5393 sibcall_failure = 1;
5395 /* Don't allow anything left on stack from computation
5396 of argument to alloca. */
5397 if (flags & ECF_MAY_BE_ALLOCA)
5398 do_pending_stack_adjust ();
5400 if (arg->value == arg->stack)
5401 /* If the value is already in the stack slot, we are done. */
5403 else if (arg->mode != BLKmode)
5405 int size;
5406 unsigned int parm_align;
5408 /* Argument is a scalar, not entirely passed in registers.
5409 (If part is passed in registers, arg->partial says how much
5410 and emit_push_insn will take care of putting it there.)
5412 Push it, and if its size is less than the
5413 amount of space allocated to it,
5414 also bump stack pointer by the additional space.
5415 Note that in C the default argument promotions
5416 will prevent such mismatches. */
5418 size = GET_MODE_SIZE (arg->mode);
5419 /* Compute how much space the push instruction will push.
5420 On many machines, pushing a byte will advance the stack
5421 pointer by a halfword. */
5422 #ifdef PUSH_ROUNDING
5423 size = PUSH_ROUNDING (size);
5424 #endif
5425 used = size;
5427 /* Compute how much space the argument should get:
5428 round up to a multiple of the alignment for arguments. */
5429 if (none != FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)))
5430 used = (((size + PARM_BOUNDARY / BITS_PER_UNIT - 1)
5431 / (PARM_BOUNDARY / BITS_PER_UNIT))
5432 * (PARM_BOUNDARY / BITS_PER_UNIT));
5434 /* Compute the alignment of the pushed argument. */
5435 parm_align = arg->locate.boundary;
5436 if (FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)) == downward)
5438 int pad = used - size;
5439 if (pad)
5441 unsigned int pad_align = least_bit_hwi (pad) * BITS_PER_UNIT;
5442 parm_align = MIN (parm_align, pad_align);
5446 /* This isn't already where we want it on the stack, so put it there.
5447 This can either be done with push or copy insns. */
5448 if (!emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), NULL_RTX,
5449 parm_align, partial, reg, used - size, argblock,
5450 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space,
5451 ARGS_SIZE_RTX (arg->locate.alignment_pad), true))
5452 sibcall_failure = 1;
5454 /* Unless this is a partially-in-register argument, the argument is now
5455 in the stack. */
5456 if (partial == 0)
5457 arg->value = arg->stack;
5459 else
5461 /* BLKmode, at least partly to be pushed. */
5463 unsigned int parm_align;
5464 int excess;
5465 rtx size_rtx;
5467 /* Pushing a nonscalar.
5468 If part is passed in registers, PARTIAL says how much
5469 and emit_push_insn will take care of putting it there. */
5471 /* Round its size up to a multiple
5472 of the allocation unit for arguments. */
5474 if (arg->locate.size.var != 0)
5476 excess = 0;
5477 size_rtx = ARGS_SIZE_RTX (arg->locate.size);
5479 else
5481 /* PUSH_ROUNDING has no effect on us, because emit_push_insn
5482 for BLKmode is careful to avoid it. */
5483 excess = (arg->locate.size.constant
5484 - int_size_in_bytes (TREE_TYPE (pval))
5485 + partial);
5486 size_rtx = expand_expr (size_in_bytes (TREE_TYPE (pval)),
5487 NULL_RTX, TYPE_MODE (sizetype),
5488 EXPAND_NORMAL);
5491 parm_align = arg->locate.boundary;
5493 /* When an argument is padded down, the block is aligned to
5494 PARM_BOUNDARY, but the actual argument isn't. */
5495 if (FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)) == downward)
5497 if (arg->locate.size.var)
5498 parm_align = BITS_PER_UNIT;
5499 else if (excess)
5501 unsigned int excess_align = least_bit_hwi (excess) * BITS_PER_UNIT;
5502 parm_align = MIN (parm_align, excess_align);
5506 if ((flags & ECF_SIBCALL) && MEM_P (arg->value))
5508 /* emit_push_insn might not work properly if arg->value and
5509 argblock + arg->locate.offset areas overlap. */
5510 rtx x = arg->value;
5511 int i = 0;
5513 if (XEXP (x, 0) == crtl->args.internal_arg_pointer
5514 || (GET_CODE (XEXP (x, 0)) == PLUS
5515 && XEXP (XEXP (x, 0), 0) ==
5516 crtl->args.internal_arg_pointer
5517 && CONST_INT_P (XEXP (XEXP (x, 0), 1))))
5519 if (XEXP (x, 0) != crtl->args.internal_arg_pointer)
5520 i = INTVAL (XEXP (XEXP (x, 0), 1));
5522 /* arg.locate doesn't contain the pretend_args_size offset,
5523 it's part of argblock. Ensure we don't count it in I. */
5524 if (STACK_GROWS_DOWNWARD)
5525 i -= crtl->args.pretend_args_size;
5526 else
5527 i += crtl->args.pretend_args_size;
5529 /* expand_call should ensure this. */
5530 gcc_assert (!arg->locate.offset.var
5531 && arg->locate.size.var == 0
5532 && CONST_INT_P (size_rtx));
5534 if (arg->locate.offset.constant > i)
5536 if (arg->locate.offset.constant < i + INTVAL (size_rtx))
5537 sibcall_failure = 1;
5539 else if (arg->locate.offset.constant < i)
5541 /* Use arg->locate.size.constant instead of size_rtx
5542 because we only care about the part of the argument
5543 on the stack. */
5544 if (i < (arg->locate.offset.constant
5545 + arg->locate.size.constant))
5546 sibcall_failure = 1;
5548 else
5550 /* Even though they appear to be at the same location,
5551 if part of the outgoing argument is in registers,
5552 they aren't really at the same location. Check for
5553 this by making sure that the incoming size is the
5554 same as the outgoing size. */
5555 if (arg->locate.size.constant != INTVAL (size_rtx))
5556 sibcall_failure = 1;
5561 emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), size_rtx,
5562 parm_align, partial, reg, excess, argblock,
5563 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space,
5564 ARGS_SIZE_RTX (arg->locate.alignment_pad), false);
5566 /* Unless this is a partially-in-register argument, the argument is now
5567 in the stack.
5569 ??? Unlike the case above, in which we want the actual
5570 address of the data, so that we can load it directly into a
5571 register, here we want the address of the stack slot, so that
5572 it's properly aligned for word-by-word copying or something
5573 like that. It's not clear that this is always correct. */
5574 if (partial == 0)
5575 arg->value = arg->stack_slot;
5578 if (arg->reg && GET_CODE (arg->reg) == PARALLEL)
5580 tree type = TREE_TYPE (arg->tree_value);
5581 arg->parallel_value
5582 = emit_group_load_into_temps (arg->reg, arg->value, type,
5583 int_size_in_bytes (type));
5586 /* Mark all slots this store used. */
5587 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL)
5588 && argblock && ! variable_size && arg->stack)
5589 for (i = lower_bound; i < upper_bound; i++)
5590 stack_usage_map[i] = 1;
5592 /* Once we have pushed something, pops can't safely
5593 be deferred during the rest of the arguments. */
5594 NO_DEFER_POP;
5596 /* Free any temporary slots made in processing this argument. */
5597 pop_temp_slots ();
5599 return sibcall_failure;
5602 /* Nonzero if we do not know how to pass TYPE solely in registers. */
5604 bool
5605 must_pass_in_stack_var_size (machine_mode mode ATTRIBUTE_UNUSED,
5606 const_tree type)
5608 if (!type)
5609 return false;
5611 /* If the type has variable size... */
5612 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
5613 return true;
5615 /* If the type is marked as addressable (it is required
5616 to be constructed into the stack)... */
5617 if (TREE_ADDRESSABLE (type))
5618 return true;
5620 return false;
5623 /* Another version of the TARGET_MUST_PASS_IN_STACK hook. This one
5624 takes trailing padding of a structure into account. */
5625 /* ??? Should be able to merge these two by examining BLOCK_REG_PADDING. */
5627 bool
5628 must_pass_in_stack_var_size_or_pad (machine_mode mode, const_tree type)
5630 if (!type)
5631 return false;
5633 /* If the type has variable size... */
5634 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
5635 return true;
5637 /* If the type is marked as addressable (it is required
5638 to be constructed into the stack)... */
5639 if (TREE_ADDRESSABLE (type))
5640 return true;
5642 /* If the padding and mode of the type is such that a copy into
5643 a register would put it into the wrong part of the register. */
5644 if (mode == BLKmode
5645 && int_size_in_bytes (type) % (PARM_BOUNDARY / BITS_PER_UNIT)
5646 && (FUNCTION_ARG_PADDING (mode, type)
5647 == (BYTES_BIG_ENDIAN ? upward : downward)))
5648 return true;
5650 return false;