PR rtl-optimization/82913
[official-gcc.git] / gcc / calls.c
blob3730f43c7a9640de804683aea6cb72383b61f7ec
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"
55 #include "stringpool.h"
56 #include "attribs.h"
58 /* Like PREFERRED_STACK_BOUNDARY but in units of bytes, not bits. */
59 #define STACK_BYTES (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT)
61 /* Data structure and subroutines used within expand_call. */
63 struct arg_data
65 /* Tree node for this argument. */
66 tree tree_value;
67 /* Mode for value; TYPE_MODE unless promoted. */
68 machine_mode mode;
69 /* Current RTL value for argument, or 0 if it isn't precomputed. */
70 rtx value;
71 /* Initially-compute RTL value for argument; only for const functions. */
72 rtx initial_value;
73 /* Register to pass this argument in, 0 if passed on stack, or an
74 PARALLEL if the arg is to be copied into multiple non-contiguous
75 registers. */
76 rtx reg;
77 /* Register to pass this argument in when generating tail call sequence.
78 This is not the same register as for normal calls on machines with
79 register windows. */
80 rtx tail_call_reg;
81 /* If REG is a PARALLEL, this is a copy of VALUE pulled into the correct
82 form for emit_group_move. */
83 rtx parallel_value;
84 /* If value is passed in neither reg nor stack, this field holds a number
85 of a special slot to be used. */
86 rtx special_slot;
87 /* For pointer bounds hold an index of parm bounds are bound to. -1 if
88 there is no such pointer. */
89 int pointer_arg;
90 /* If pointer_arg refers a structure, then pointer_offset holds an offset
91 of a pointer in this structure. */
92 int pointer_offset;
93 /* If REG was promoted from the actual mode of the argument expression,
94 indicates whether the promotion is sign- or zero-extended. */
95 int unsignedp;
96 /* Number of bytes to put in registers. 0 means put the whole arg
97 in registers. Also 0 if not passed in registers. */
98 int partial;
99 /* Nonzero if argument must be passed on stack.
100 Note that some arguments may be passed on the stack
101 even though pass_on_stack is zero, just because FUNCTION_ARG says so.
102 pass_on_stack identifies arguments that *cannot* go in registers. */
103 int pass_on_stack;
104 /* Some fields packaged up for locate_and_pad_parm. */
105 struct locate_and_pad_arg_data locate;
106 /* Location on the stack at which parameter should be stored. The store
107 has already been done if STACK == VALUE. */
108 rtx stack;
109 /* Location on the stack of the start of this argument slot. This can
110 differ from STACK if this arg pads downward. This location is known
111 to be aligned to TARGET_FUNCTION_ARG_BOUNDARY. */
112 rtx stack_slot;
113 /* Place that this stack area has been saved, if needed. */
114 rtx save_area;
115 /* If an argument's alignment does not permit direct copying into registers,
116 copy in smaller-sized pieces into pseudos. These are stored in a
117 block pointed to by this field. The next field says how many
118 word-sized pseudos we made. */
119 rtx *aligned_regs;
120 int n_aligned_regs;
123 /* A vector of one char per byte of stack space. A byte if nonzero if
124 the corresponding stack location has been used.
125 This vector is used to prevent a function call within an argument from
126 clobbering any stack already set up. */
127 static char *stack_usage_map;
129 /* Size of STACK_USAGE_MAP. */
130 static int highest_outgoing_arg_in_use;
132 /* A bitmap of virtual-incoming stack space. Bit is set if the corresponding
133 stack location's tail call argument has been already stored into the stack.
134 This bitmap is used to prevent sibling call optimization if function tries
135 to use parent's incoming argument slots when they have been already
136 overwritten with tail call arguments. */
137 static sbitmap stored_args_map;
139 /* stack_arg_under_construction is nonzero when an argument may be
140 initialized with a constructor call (including a C function that
141 returns a BLKmode struct) and expand_call must take special action
142 to make sure the object being constructed does not overlap the
143 argument list for the constructor call. */
144 static int stack_arg_under_construction;
146 static void emit_call_1 (rtx, tree, tree, tree, HOST_WIDE_INT, HOST_WIDE_INT,
147 HOST_WIDE_INT, rtx, rtx, int, rtx, int,
148 cumulative_args_t);
149 static void precompute_register_parameters (int, struct arg_data *, int *);
150 static void store_bounds (struct arg_data *, struct arg_data *);
151 static int store_one_arg (struct arg_data *, rtx, int, int, int);
152 static void store_unaligned_arguments_into_pseudos (struct arg_data *, int);
153 static int finalize_must_preallocate (int, int, struct arg_data *,
154 struct args_size *);
155 static void precompute_arguments (int, struct arg_data *);
156 static int compute_argument_block_size (int, struct args_size *, tree, tree, int);
157 static void initialize_argument_information (int, struct arg_data *,
158 struct args_size *, int,
159 tree, tree,
160 tree, tree, cumulative_args_t, int,
161 rtx *, int *, int *, int *,
162 bool *, bool);
163 static void compute_argument_addresses (struct arg_data *, rtx, int);
164 static rtx rtx_for_function_call (tree, tree);
165 static void load_register_parameters (struct arg_data *, int, rtx *, int,
166 int, int *);
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 if (GET_MODE (funexp) != Pmode)
210 funexp = convert_memory_address (Pmode, funexp);
212 /* Avoid long live ranges around function calls. */
213 funexp = copy_to_mode_reg (Pmode, funexp);
215 if (REG_P (chain))
216 emit_insn (gen_rtx_CLOBBER (VOIDmode, chain));
218 /* Emit the runtime identification pattern. */
219 rtx mask = gen_rtx_AND (Pmode, funexp, GEN_INT (bit_val));
220 emit_cmp_and_jump_insns (mask, const0_rtx, EQ, NULL_RTX, Pmode, 1,
221 call_lab);
223 /* Statically predict the branch to very likely taken. */
224 rtx_insn *insn = get_last_insn ();
225 if (JUMP_P (insn))
226 predict_insn_def (insn, PRED_BUILTIN_EXPECT, TAKEN);
228 /* Load the descriptor. */
229 rtx mem = gen_rtx_MEM (ptr_mode,
230 plus_constant (Pmode, funexp, - bit_val));
231 MEM_NOTRAP_P (mem) = 1;
232 mem = convert_memory_address (Pmode, mem);
233 emit_move_insn (chain, mem);
235 mem = gen_rtx_MEM (ptr_mode,
236 plus_constant (Pmode, funexp,
237 POINTER_SIZE / BITS_PER_UNIT
238 - bit_val));
239 MEM_NOTRAP_P (mem) = 1;
240 mem = convert_memory_address (Pmode, mem);
241 emit_move_insn (funexp, mem);
243 emit_label (call_lab);
245 if (REG_P (chain))
247 use_reg (call_fusage, chain);
248 STATIC_CHAIN_REG_P (chain) = 1;
251 /* Make sure we're not going to be overwritten below. */
252 gcc_assert (!static_chain_value);
255 /* If we are using registers for parameters, force the
256 function address into a register now. */
257 funexp = ((reg_parm_seen
258 && targetm.small_register_classes_for_mode_p (FUNCTION_MODE))
259 ? force_not_mem (memory_address (FUNCTION_MODE, funexp))
260 : memory_address (FUNCTION_MODE, funexp));
262 else
264 /* funexp could be a SYMBOL_REF represents a function pointer which is
265 of ptr_mode. In this case, it should be converted into address mode
266 to be a valid address for memory rtx pattern. See PR 64971. */
267 if (GET_MODE (funexp) != Pmode)
268 funexp = convert_memory_address (Pmode, funexp);
270 if (!(flags & ECF_SIBCALL))
272 if (!NO_FUNCTION_CSE && optimize && ! flag_no_function_cse)
273 funexp = force_reg (Pmode, funexp);
277 if (static_chain_value != 0
278 && (TREE_CODE (fndecl_or_type) != FUNCTION_DECL
279 || DECL_STATIC_CHAIN (fndecl_or_type)))
281 rtx chain;
283 chain = targetm.calls.static_chain (fndecl_or_type, false);
284 static_chain_value = convert_memory_address (Pmode, static_chain_value);
286 emit_move_insn (chain, static_chain_value);
287 if (REG_P (chain))
289 use_reg (call_fusage, chain);
290 STATIC_CHAIN_REG_P (chain) = 1;
294 return funexp;
297 /* Generate instructions to call function FUNEXP,
298 and optionally pop the results.
299 The CALL_INSN is the first insn generated.
301 FNDECL is the declaration node of the function. This is given to the
302 hook TARGET_RETURN_POPS_ARGS to determine whether this function pops
303 its own args.
305 FUNTYPE is the data type of the function. This is given to the hook
306 TARGET_RETURN_POPS_ARGS to determine whether this function pops its
307 own args. We used to allow an identifier for library functions, but
308 that doesn't work when the return type is an aggregate type and the
309 calling convention says that the pointer to this aggregate is to be
310 popped by the callee.
312 STACK_SIZE is the number of bytes of arguments on the stack,
313 ROUNDED_STACK_SIZE is that number rounded up to
314 PREFERRED_STACK_BOUNDARY; zero if the size is variable. This is
315 both to put into the call insn and to generate explicit popping
316 code if necessary.
318 STRUCT_VALUE_SIZE is the number of bytes wanted in a structure value.
319 It is zero if this call doesn't want a structure value.
321 NEXT_ARG_REG is the rtx that results from executing
322 targetm.calls.function_arg (&args_so_far, VOIDmode, void_type_node, true)
323 just after all the args have had their registers assigned.
324 This could be whatever you like, but normally it is the first
325 arg-register beyond those used for args in this call,
326 or 0 if all the arg-registers are used in this call.
327 It is passed on to `gen_call' so you can put this info in the call insn.
329 VALREG is a hard register in which a value is returned,
330 or 0 if the call does not return a value.
332 OLD_INHIBIT_DEFER_POP is the value that `inhibit_defer_pop' had before
333 the args to this call were processed.
334 We restore `inhibit_defer_pop' to that value.
336 CALL_FUSAGE is either empty or an EXPR_LIST of USE expressions that
337 denote registers used by the called function. */
339 static void
340 emit_call_1 (rtx funexp, tree fntree ATTRIBUTE_UNUSED, tree fndecl ATTRIBUTE_UNUSED,
341 tree funtype ATTRIBUTE_UNUSED,
342 HOST_WIDE_INT stack_size ATTRIBUTE_UNUSED,
343 HOST_WIDE_INT rounded_stack_size,
344 HOST_WIDE_INT struct_value_size ATTRIBUTE_UNUSED,
345 rtx next_arg_reg ATTRIBUTE_UNUSED, rtx valreg,
346 int old_inhibit_defer_pop, rtx call_fusage, int ecf_flags,
347 cumulative_args_t args_so_far ATTRIBUTE_UNUSED)
349 rtx rounded_stack_size_rtx = GEN_INT (rounded_stack_size);
350 rtx call, funmem, pat;
351 int already_popped = 0;
352 HOST_WIDE_INT n_popped = 0;
354 /* Sibling call patterns never pop arguments (no sibcall(_value)_pop
355 patterns exist). Any popping that the callee does on return will
356 be from our caller's frame rather than ours. */
357 if (!(ecf_flags & ECF_SIBCALL))
359 n_popped += targetm.calls.return_pops_args (fndecl, funtype, stack_size);
361 #ifdef CALL_POPS_ARGS
362 n_popped += CALL_POPS_ARGS (*get_cumulative_args (args_so_far));
363 #endif
366 /* Ensure address is valid. SYMBOL_REF is already valid, so no need,
367 and we don't want to load it into a register as an optimization,
368 because prepare_call_address already did it if it should be done. */
369 if (GET_CODE (funexp) != SYMBOL_REF)
370 funexp = memory_address (FUNCTION_MODE, funexp);
372 funmem = gen_rtx_MEM (FUNCTION_MODE, funexp);
373 if (fndecl && TREE_CODE (fndecl) == FUNCTION_DECL)
375 tree t = fndecl;
377 /* Although a built-in FUNCTION_DECL and its non-__builtin
378 counterpart compare equal and get a shared mem_attrs, they
379 produce different dump output in compare-debug compilations,
380 if an entry gets garbage collected in one compilation, then
381 adds a different (but equivalent) entry, while the other
382 doesn't run the garbage collector at the same spot and then
383 shares the mem_attr with the equivalent entry. */
384 if (DECL_BUILT_IN_CLASS (t) == BUILT_IN_NORMAL)
386 tree t2 = builtin_decl_explicit (DECL_FUNCTION_CODE (t));
387 if (t2)
388 t = t2;
391 set_mem_expr (funmem, t);
393 else if (fntree)
394 set_mem_expr (funmem, build_simple_mem_ref (CALL_EXPR_FN (fntree)));
396 if (ecf_flags & ECF_SIBCALL)
398 if (valreg)
399 pat = targetm.gen_sibcall_value (valreg, funmem,
400 rounded_stack_size_rtx,
401 next_arg_reg, NULL_RTX);
402 else
403 pat = targetm.gen_sibcall (funmem, rounded_stack_size_rtx,
404 next_arg_reg, GEN_INT (struct_value_size));
406 /* If the target has "call" or "call_value" insns, then prefer them
407 if no arguments are actually popped. If the target does not have
408 "call" or "call_value" insns, then we must use the popping versions
409 even if the call has no arguments to pop. */
410 else if (n_popped > 0
411 || !(valreg
412 ? targetm.have_call_value ()
413 : targetm.have_call ()))
415 rtx n_pop = GEN_INT (n_popped);
417 /* If this subroutine pops its own args, record that in the call insn
418 if possible, for the sake of frame pointer elimination. */
420 if (valreg)
421 pat = targetm.gen_call_value_pop (valreg, funmem,
422 rounded_stack_size_rtx,
423 next_arg_reg, n_pop);
424 else
425 pat = targetm.gen_call_pop (funmem, rounded_stack_size_rtx,
426 next_arg_reg, n_pop);
428 already_popped = 1;
430 else
432 if (valreg)
433 pat = targetm.gen_call_value (valreg, funmem, rounded_stack_size_rtx,
434 next_arg_reg, NULL_RTX);
435 else
436 pat = targetm.gen_call (funmem, rounded_stack_size_rtx, next_arg_reg,
437 GEN_INT (struct_value_size));
439 emit_insn (pat);
441 /* Find the call we just emitted. */
442 rtx_call_insn *call_insn = last_call_insn ();
444 /* Some target create a fresh MEM instead of reusing the one provided
445 above. Set its MEM_EXPR. */
446 call = get_call_rtx_from (call_insn);
447 if (call
448 && MEM_EXPR (XEXP (call, 0)) == NULL_TREE
449 && MEM_EXPR (funmem) != NULL_TREE)
450 set_mem_expr (XEXP (call, 0), MEM_EXPR (funmem));
452 /* Mark instrumented calls. */
453 if (call && fntree)
454 CALL_EXPR_WITH_BOUNDS_P (call) = CALL_WITH_BOUNDS_P (fntree);
456 /* Put the register usage information there. */
457 add_function_usage_to (call_insn, call_fusage);
459 /* If this is a const call, then set the insn's unchanging bit. */
460 if (ecf_flags & ECF_CONST)
461 RTL_CONST_CALL_P (call_insn) = 1;
463 /* If this is a pure call, then set the insn's unchanging bit. */
464 if (ecf_flags & ECF_PURE)
465 RTL_PURE_CALL_P (call_insn) = 1;
467 /* If this is a const call, then set the insn's unchanging bit. */
468 if (ecf_flags & ECF_LOOPING_CONST_OR_PURE)
469 RTL_LOOPING_CONST_OR_PURE_CALL_P (call_insn) = 1;
471 /* Create a nothrow REG_EH_REGION note, if needed. */
472 make_reg_eh_region_note (call_insn, ecf_flags, 0);
474 if (ecf_flags & ECF_NORETURN)
475 add_reg_note (call_insn, REG_NORETURN, const0_rtx);
477 if (ecf_flags & ECF_RETURNS_TWICE)
479 add_reg_note (call_insn, REG_SETJMP, const0_rtx);
480 cfun->calls_setjmp = 1;
483 SIBLING_CALL_P (call_insn) = ((ecf_flags & ECF_SIBCALL) != 0);
485 /* Restore this now, so that we do defer pops for this call's args
486 if the context of the call as a whole permits. */
487 inhibit_defer_pop = old_inhibit_defer_pop;
489 if (n_popped > 0)
491 if (!already_popped)
492 CALL_INSN_FUNCTION_USAGE (call_insn)
493 = gen_rtx_EXPR_LIST (VOIDmode,
494 gen_rtx_CLOBBER (VOIDmode, stack_pointer_rtx),
495 CALL_INSN_FUNCTION_USAGE (call_insn));
496 rounded_stack_size -= n_popped;
497 rounded_stack_size_rtx = GEN_INT (rounded_stack_size);
498 stack_pointer_delta -= n_popped;
500 add_reg_note (call_insn, REG_ARGS_SIZE, GEN_INT (stack_pointer_delta));
502 /* If popup is needed, stack realign must use DRAP */
503 if (SUPPORTS_STACK_ALIGNMENT)
504 crtl->need_drap = true;
506 /* For noreturn calls when not accumulating outgoing args force
507 REG_ARGS_SIZE note to prevent crossjumping of calls with different
508 args sizes. */
509 else if (!ACCUMULATE_OUTGOING_ARGS && (ecf_flags & ECF_NORETURN) != 0)
510 add_reg_note (call_insn, REG_ARGS_SIZE, GEN_INT (stack_pointer_delta));
512 if (!ACCUMULATE_OUTGOING_ARGS)
514 /* If returning from the subroutine does not automatically pop the args,
515 we need an instruction to pop them sooner or later.
516 Perhaps do it now; perhaps just record how much space to pop later.
518 If returning from the subroutine does pop the args, indicate that the
519 stack pointer will be changed. */
521 if (rounded_stack_size != 0)
523 if (ecf_flags & ECF_NORETURN)
524 /* Just pretend we did the pop. */
525 stack_pointer_delta -= rounded_stack_size;
526 else if (flag_defer_pop && inhibit_defer_pop == 0
527 && ! (ecf_flags & (ECF_CONST | ECF_PURE)))
528 pending_stack_adjust += rounded_stack_size;
529 else
530 adjust_stack (rounded_stack_size_rtx);
533 /* When we accumulate outgoing args, we must avoid any stack manipulations.
534 Restore the stack pointer to its original value now. Usually
535 ACCUMULATE_OUTGOING_ARGS targets don't get here, but there are exceptions.
536 On i386 ACCUMULATE_OUTGOING_ARGS can be enabled on demand, and
537 popping variants of functions exist as well.
539 ??? We may optimize similar to defer_pop above, but it is
540 probably not worthwhile.
542 ??? It will be worthwhile to enable combine_stack_adjustments even for
543 such machines. */
544 else if (n_popped)
545 anti_adjust_stack (GEN_INT (n_popped));
548 /* Determine if the function identified by FNDECL is one with
549 special properties we wish to know about. Modify FLAGS accordingly.
551 For example, if the function might return more than one time (setjmp), then
552 set ECF_RETURNS_TWICE.
554 Set ECF_MAY_BE_ALLOCA for any memory allocation function that might allocate
555 space from the stack such as alloca. */
557 static int
558 special_function_p (const_tree fndecl, int flags)
560 tree name_decl = DECL_NAME (fndecl);
562 /* For instrumentation clones we want to derive flags
563 from the original name. */
564 if (cgraph_node::get (fndecl)
565 && cgraph_node::get (fndecl)->instrumentation_clone)
566 name_decl = DECL_NAME (cgraph_node::get (fndecl)->orig_decl);
568 if (fndecl && name_decl
569 && IDENTIFIER_LENGTH (name_decl) <= 11
570 /* Exclude functions not at the file scope, or not `extern',
571 since they are not the magic functions we would otherwise
572 think they are.
573 FIXME: this should be handled with attributes, not with this
574 hacky imitation of DECL_ASSEMBLER_NAME. It's (also) wrong
575 because you can declare fork() inside a function if you
576 wish. */
577 && (DECL_CONTEXT (fndecl) == NULL_TREE
578 || TREE_CODE (DECL_CONTEXT (fndecl)) == TRANSLATION_UNIT_DECL)
579 && TREE_PUBLIC (fndecl))
581 const char *name = IDENTIFIER_POINTER (name_decl);
582 const char *tname = name;
584 /* We assume that alloca will always be called by name. It
585 makes no sense to pass it as a pointer-to-function to
586 anything that does not understand its behavior. */
587 if (IDENTIFIER_LENGTH (name_decl) == 6
588 && name[0] == 'a'
589 && ! strcmp (name, "alloca"))
590 flags |= ECF_MAY_BE_ALLOCA;
592 /* Disregard prefix _ or __. */
593 if (name[0] == '_')
595 if (name[1] == '_')
596 tname += 2;
597 else
598 tname += 1;
601 /* ECF_RETURNS_TWICE is safe even for -ffreestanding. */
602 if (! strcmp (tname, "setjmp")
603 || ! strcmp (tname, "sigsetjmp")
604 || ! strcmp (name, "savectx")
605 || ! strcmp (name, "vfork")
606 || ! strcmp (name, "getcontext"))
607 flags |= ECF_RETURNS_TWICE;
610 if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL
611 && ALLOCA_FUNCTION_CODE_P (DECL_FUNCTION_CODE (fndecl)))
612 flags |= ECF_MAY_BE_ALLOCA;
614 return flags;
617 /* Similar to special_function_p; return a set of ERF_ flags for the
618 function FNDECL. */
619 static int
620 decl_return_flags (tree fndecl)
622 tree attr;
623 tree type = TREE_TYPE (fndecl);
624 if (!type)
625 return 0;
627 attr = lookup_attribute ("fn spec", TYPE_ATTRIBUTES (type));
628 if (!attr)
629 return 0;
631 attr = TREE_VALUE (TREE_VALUE (attr));
632 if (!attr || TREE_STRING_LENGTH (attr) < 1)
633 return 0;
635 switch (TREE_STRING_POINTER (attr)[0])
637 case '1':
638 case '2':
639 case '3':
640 case '4':
641 return ERF_RETURNS_ARG | (TREE_STRING_POINTER (attr)[0] - '1');
643 case 'm':
644 return ERF_NOALIAS;
646 case '.':
647 default:
648 return 0;
652 /* Return nonzero when FNDECL represents a call to setjmp. */
655 setjmp_call_p (const_tree fndecl)
657 if (DECL_IS_RETURNS_TWICE (fndecl))
658 return ECF_RETURNS_TWICE;
659 return special_function_p (fndecl, 0) & ECF_RETURNS_TWICE;
663 /* Return true if STMT may be an alloca call. */
665 bool
666 gimple_maybe_alloca_call_p (const gimple *stmt)
668 tree fndecl;
670 if (!is_gimple_call (stmt))
671 return false;
673 fndecl = gimple_call_fndecl (stmt);
674 if (fndecl && (special_function_p (fndecl, 0) & ECF_MAY_BE_ALLOCA))
675 return true;
677 return false;
680 /* Return true if STMT is a builtin alloca call. */
682 bool
683 gimple_alloca_call_p (const gimple *stmt)
685 tree fndecl;
687 if (!is_gimple_call (stmt))
688 return false;
690 fndecl = gimple_call_fndecl (stmt);
691 if (fndecl && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
692 switch (DECL_FUNCTION_CODE (fndecl))
694 CASE_BUILT_IN_ALLOCA:
695 return true;
696 default:
697 break;
700 return false;
703 /* Return true when exp contains a builtin alloca call. */
705 bool
706 alloca_call_p (const_tree exp)
708 tree fndecl;
709 if (TREE_CODE (exp) == CALL_EXPR
710 && (fndecl = get_callee_fndecl (exp))
711 && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
712 switch (DECL_FUNCTION_CODE (fndecl))
714 CASE_BUILT_IN_ALLOCA:
715 return true;
716 default:
717 break;
720 return false;
723 /* Return TRUE if FNDECL is either a TM builtin or a TM cloned
724 function. Return FALSE otherwise. */
726 static bool
727 is_tm_builtin (const_tree fndecl)
729 if (fndecl == NULL)
730 return false;
732 if (decl_is_tm_clone (fndecl))
733 return true;
735 if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
737 switch (DECL_FUNCTION_CODE (fndecl))
739 case BUILT_IN_TM_COMMIT:
740 case BUILT_IN_TM_COMMIT_EH:
741 case BUILT_IN_TM_ABORT:
742 case BUILT_IN_TM_IRREVOCABLE:
743 case BUILT_IN_TM_GETTMCLONE_IRR:
744 case BUILT_IN_TM_MEMCPY:
745 case BUILT_IN_TM_MEMMOVE:
746 case BUILT_IN_TM_MEMSET:
747 CASE_BUILT_IN_TM_STORE (1):
748 CASE_BUILT_IN_TM_STORE (2):
749 CASE_BUILT_IN_TM_STORE (4):
750 CASE_BUILT_IN_TM_STORE (8):
751 CASE_BUILT_IN_TM_STORE (FLOAT):
752 CASE_BUILT_IN_TM_STORE (DOUBLE):
753 CASE_BUILT_IN_TM_STORE (LDOUBLE):
754 CASE_BUILT_IN_TM_STORE (M64):
755 CASE_BUILT_IN_TM_STORE (M128):
756 CASE_BUILT_IN_TM_STORE (M256):
757 CASE_BUILT_IN_TM_LOAD (1):
758 CASE_BUILT_IN_TM_LOAD (2):
759 CASE_BUILT_IN_TM_LOAD (4):
760 CASE_BUILT_IN_TM_LOAD (8):
761 CASE_BUILT_IN_TM_LOAD (FLOAT):
762 CASE_BUILT_IN_TM_LOAD (DOUBLE):
763 CASE_BUILT_IN_TM_LOAD (LDOUBLE):
764 CASE_BUILT_IN_TM_LOAD (M64):
765 CASE_BUILT_IN_TM_LOAD (M128):
766 CASE_BUILT_IN_TM_LOAD (M256):
767 case BUILT_IN_TM_LOG:
768 case BUILT_IN_TM_LOG_1:
769 case BUILT_IN_TM_LOG_2:
770 case BUILT_IN_TM_LOG_4:
771 case BUILT_IN_TM_LOG_8:
772 case BUILT_IN_TM_LOG_FLOAT:
773 case BUILT_IN_TM_LOG_DOUBLE:
774 case BUILT_IN_TM_LOG_LDOUBLE:
775 case BUILT_IN_TM_LOG_M64:
776 case BUILT_IN_TM_LOG_M128:
777 case BUILT_IN_TM_LOG_M256:
778 return true;
779 default:
780 break;
783 return false;
786 /* Detect flags (function attributes) from the function decl or type node. */
789 flags_from_decl_or_type (const_tree exp)
791 int flags = 0;
793 if (DECL_P (exp))
795 /* The function exp may have the `malloc' attribute. */
796 if (DECL_IS_MALLOC (exp))
797 flags |= ECF_MALLOC;
799 /* The function exp may have the `returns_twice' attribute. */
800 if (DECL_IS_RETURNS_TWICE (exp))
801 flags |= ECF_RETURNS_TWICE;
803 /* Process the pure and const attributes. */
804 if (TREE_READONLY (exp))
805 flags |= ECF_CONST;
806 if (DECL_PURE_P (exp))
807 flags |= ECF_PURE;
808 if (DECL_LOOPING_CONST_OR_PURE_P (exp))
809 flags |= ECF_LOOPING_CONST_OR_PURE;
811 if (DECL_IS_NOVOPS (exp))
812 flags |= ECF_NOVOPS;
813 if (lookup_attribute ("leaf", DECL_ATTRIBUTES (exp)))
814 flags |= ECF_LEAF;
815 if (lookup_attribute ("cold", DECL_ATTRIBUTES (exp)))
816 flags |= ECF_COLD;
818 if (TREE_NOTHROW (exp))
819 flags |= ECF_NOTHROW;
821 if (flag_tm)
823 if (is_tm_builtin (exp))
824 flags |= ECF_TM_BUILTIN;
825 else if ((flags & (ECF_CONST|ECF_NOVOPS)) != 0
826 || lookup_attribute ("transaction_pure",
827 TYPE_ATTRIBUTES (TREE_TYPE (exp))))
828 flags |= ECF_TM_PURE;
831 flags = special_function_p (exp, flags);
833 else if (TYPE_P (exp))
835 if (TYPE_READONLY (exp))
836 flags |= ECF_CONST;
838 if (flag_tm
839 && ((flags & ECF_CONST) != 0
840 || lookup_attribute ("transaction_pure", TYPE_ATTRIBUTES (exp))))
841 flags |= ECF_TM_PURE;
843 else
844 gcc_unreachable ();
846 if (TREE_THIS_VOLATILE (exp))
848 flags |= ECF_NORETURN;
849 if (flags & (ECF_CONST|ECF_PURE))
850 flags |= ECF_LOOPING_CONST_OR_PURE;
853 return flags;
856 /* Detect flags from a CALL_EXPR. */
859 call_expr_flags (const_tree t)
861 int flags;
862 tree decl = get_callee_fndecl (t);
864 if (decl)
865 flags = flags_from_decl_or_type (decl);
866 else if (CALL_EXPR_FN (t) == NULL_TREE)
867 flags = internal_fn_flags (CALL_EXPR_IFN (t));
868 else
870 tree type = TREE_TYPE (CALL_EXPR_FN (t));
871 if (type && TREE_CODE (type) == POINTER_TYPE)
872 flags = flags_from_decl_or_type (TREE_TYPE (type));
873 else
874 flags = 0;
875 if (CALL_EXPR_BY_DESCRIPTOR (t))
876 flags |= ECF_BY_DESCRIPTOR;
879 return flags;
882 /* Return true if TYPE should be passed by invisible reference. */
884 bool
885 pass_by_reference (CUMULATIVE_ARGS *ca, machine_mode mode,
886 tree type, bool named_arg)
888 if (type)
890 /* If this type contains non-trivial constructors, then it is
891 forbidden for the middle-end to create any new copies. */
892 if (TREE_ADDRESSABLE (type))
893 return true;
895 /* GCC post 3.4 passes *all* variable sized types by reference. */
896 if (!TYPE_SIZE (type) || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
897 return true;
899 /* If a record type should be passed the same as its first (and only)
900 member, use the type and mode of that member. */
901 if (TREE_CODE (type) == RECORD_TYPE && TYPE_TRANSPARENT_AGGR (type))
903 type = TREE_TYPE (first_field (type));
904 mode = TYPE_MODE (type);
908 return targetm.calls.pass_by_reference (pack_cumulative_args (ca), mode,
909 type, named_arg);
912 /* Return true if TYPE, which is passed by reference, should be callee
913 copied instead of caller copied. */
915 bool
916 reference_callee_copied (CUMULATIVE_ARGS *ca, machine_mode mode,
917 tree type, bool named_arg)
919 if (type && TREE_ADDRESSABLE (type))
920 return false;
921 return targetm.calls.callee_copies (pack_cumulative_args (ca), mode, type,
922 named_arg);
926 /* Precompute all register parameters as described by ARGS, storing values
927 into fields within the ARGS array.
929 NUM_ACTUALS indicates the total number elements in the ARGS array.
931 Set REG_PARM_SEEN if we encounter a register parameter. */
933 static void
934 precompute_register_parameters (int num_actuals, struct arg_data *args,
935 int *reg_parm_seen)
937 int i;
939 *reg_parm_seen = 0;
941 for (i = 0; i < num_actuals; i++)
942 if (args[i].reg != 0 && ! args[i].pass_on_stack)
944 *reg_parm_seen = 1;
946 if (args[i].value == 0)
948 push_temp_slots ();
949 args[i].value = expand_normal (args[i].tree_value);
950 preserve_temp_slots (args[i].value);
951 pop_temp_slots ();
954 /* If we are to promote the function arg to a wider mode,
955 do it now. */
957 if (args[i].mode != TYPE_MODE (TREE_TYPE (args[i].tree_value)))
958 args[i].value
959 = convert_modes (args[i].mode,
960 TYPE_MODE (TREE_TYPE (args[i].tree_value)),
961 args[i].value, args[i].unsignedp);
963 /* If the value is a non-legitimate constant, force it into a
964 pseudo now. TLS symbols sometimes need a call to resolve. */
965 if (CONSTANT_P (args[i].value)
966 && !targetm.legitimate_constant_p (args[i].mode, args[i].value))
967 args[i].value = force_reg (args[i].mode, args[i].value);
969 /* If we're going to have to load the value by parts, pull the
970 parts into pseudos. The part extraction process can involve
971 non-trivial computation. */
972 if (GET_CODE (args[i].reg) == PARALLEL)
974 tree type = TREE_TYPE (args[i].tree_value);
975 args[i].parallel_value
976 = emit_group_load_into_temps (args[i].reg, args[i].value,
977 type, int_size_in_bytes (type));
980 /* If the value is expensive, and we are inside an appropriately
981 short loop, put the value into a pseudo and then put the pseudo
982 into the hard reg.
984 For small register classes, also do this if this call uses
985 register parameters. This is to avoid reload conflicts while
986 loading the parameters registers. */
988 else if ((! (REG_P (args[i].value)
989 || (GET_CODE (args[i].value) == SUBREG
990 && REG_P (SUBREG_REG (args[i].value)))))
991 && args[i].mode != BLKmode
992 && (set_src_cost (args[i].value, args[i].mode,
993 optimize_insn_for_speed_p ())
994 > COSTS_N_INSNS (1))
995 && ((*reg_parm_seen
996 && targetm.small_register_classes_for_mode_p (args[i].mode))
997 || optimize))
998 args[i].value = copy_to_mode_reg (args[i].mode, args[i].value);
1002 #ifdef REG_PARM_STACK_SPACE
1004 /* The argument list is the property of the called routine and it
1005 may clobber it. If the fixed area has been used for previous
1006 parameters, we must save and restore it. */
1008 static rtx
1009 save_fixed_argument_area (int reg_parm_stack_space, rtx argblock, int *low_to_save, int *high_to_save)
1011 int low;
1012 int high;
1014 /* Compute the boundary of the area that needs to be saved, if any. */
1015 high = reg_parm_stack_space;
1016 if (ARGS_GROW_DOWNWARD)
1017 high += 1;
1019 if (high > highest_outgoing_arg_in_use)
1020 high = highest_outgoing_arg_in_use;
1022 for (low = 0; low < high; low++)
1023 if (stack_usage_map[low] != 0)
1025 int num_to_save;
1026 machine_mode save_mode;
1027 int delta;
1028 rtx addr;
1029 rtx stack_area;
1030 rtx save_area;
1032 while (stack_usage_map[--high] == 0)
1035 *low_to_save = low;
1036 *high_to_save = high;
1038 num_to_save = high - low + 1;
1040 /* If we don't have the required alignment, must do this
1041 in BLKmode. */
1042 scalar_int_mode imode;
1043 if (int_mode_for_size (num_to_save * BITS_PER_UNIT, 1).exists (&imode)
1044 && (low & (MIN (GET_MODE_SIZE (imode),
1045 BIGGEST_ALIGNMENT / UNITS_PER_WORD) - 1)) == 0)
1046 save_mode = imode;
1047 else
1048 save_mode = BLKmode;
1050 if (ARGS_GROW_DOWNWARD)
1051 delta = -high;
1052 else
1053 delta = low;
1055 addr = plus_constant (Pmode, argblock, delta);
1056 stack_area = gen_rtx_MEM (save_mode, memory_address (save_mode, addr));
1058 set_mem_align (stack_area, PARM_BOUNDARY);
1059 if (save_mode == BLKmode)
1061 save_area = assign_stack_temp (BLKmode, num_to_save);
1062 emit_block_move (validize_mem (save_area), stack_area,
1063 GEN_INT (num_to_save), BLOCK_OP_CALL_PARM);
1065 else
1067 save_area = gen_reg_rtx (save_mode);
1068 emit_move_insn (save_area, stack_area);
1071 return save_area;
1074 return NULL_RTX;
1077 static void
1078 restore_fixed_argument_area (rtx save_area, rtx argblock, int high_to_save, int low_to_save)
1080 machine_mode save_mode = GET_MODE (save_area);
1081 int delta;
1082 rtx addr, stack_area;
1084 if (ARGS_GROW_DOWNWARD)
1085 delta = -high_to_save;
1086 else
1087 delta = low_to_save;
1089 addr = plus_constant (Pmode, argblock, delta);
1090 stack_area = gen_rtx_MEM (save_mode, memory_address (save_mode, addr));
1091 set_mem_align (stack_area, PARM_BOUNDARY);
1093 if (save_mode != BLKmode)
1094 emit_move_insn (stack_area, save_area);
1095 else
1096 emit_block_move (stack_area, validize_mem (save_area),
1097 GEN_INT (high_to_save - low_to_save + 1),
1098 BLOCK_OP_CALL_PARM);
1100 #endif /* REG_PARM_STACK_SPACE */
1102 /* If any elements in ARGS refer to parameters that are to be passed in
1103 registers, but not in memory, and whose alignment does not permit a
1104 direct copy into registers. Copy the values into a group of pseudos
1105 which we will later copy into the appropriate hard registers.
1107 Pseudos for each unaligned argument will be stored into the array
1108 args[argnum].aligned_regs. The caller is responsible for deallocating
1109 the aligned_regs array if it is nonzero. */
1111 static void
1112 store_unaligned_arguments_into_pseudos (struct arg_data *args, int num_actuals)
1114 int i, j;
1116 for (i = 0; i < num_actuals; i++)
1117 if (args[i].reg != 0 && ! args[i].pass_on_stack
1118 && GET_CODE (args[i].reg) != PARALLEL
1119 && args[i].mode == BLKmode
1120 && MEM_P (args[i].value)
1121 && (MEM_ALIGN (args[i].value)
1122 < (unsigned int) MIN (BIGGEST_ALIGNMENT, BITS_PER_WORD)))
1124 int bytes = int_size_in_bytes (TREE_TYPE (args[i].tree_value));
1125 int endian_correction = 0;
1127 if (args[i].partial)
1129 gcc_assert (args[i].partial % UNITS_PER_WORD == 0);
1130 args[i].n_aligned_regs = args[i].partial / UNITS_PER_WORD;
1132 else
1134 args[i].n_aligned_regs
1135 = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
1138 args[i].aligned_regs = XNEWVEC (rtx, args[i].n_aligned_regs);
1140 /* Structures smaller than a word are normally aligned to the
1141 least significant byte. On a BYTES_BIG_ENDIAN machine,
1142 this means we must skip the empty high order bytes when
1143 calculating the bit offset. */
1144 if (bytes < UNITS_PER_WORD
1145 #ifdef BLOCK_REG_PADDING
1146 && (BLOCK_REG_PADDING (args[i].mode,
1147 TREE_TYPE (args[i].tree_value), 1)
1148 == PAD_DOWNWARD)
1149 #else
1150 && BYTES_BIG_ENDIAN
1151 #endif
1153 endian_correction = BITS_PER_WORD - bytes * BITS_PER_UNIT;
1155 for (j = 0; j < args[i].n_aligned_regs; j++)
1157 rtx reg = gen_reg_rtx (word_mode);
1158 rtx word = operand_subword_force (args[i].value, j, BLKmode);
1159 int bitsize = MIN (bytes * BITS_PER_UNIT, BITS_PER_WORD);
1161 args[i].aligned_regs[j] = reg;
1162 word = extract_bit_field (word, bitsize, 0, 1, NULL_RTX,
1163 word_mode, word_mode, false, NULL);
1165 /* There is no need to restrict this code to loading items
1166 in TYPE_ALIGN sized hunks. The bitfield instructions can
1167 load up entire word sized registers efficiently.
1169 ??? This may not be needed anymore.
1170 We use to emit a clobber here but that doesn't let later
1171 passes optimize the instructions we emit. By storing 0 into
1172 the register later passes know the first AND to zero out the
1173 bitfield being set in the register is unnecessary. The store
1174 of 0 will be deleted as will at least the first AND. */
1176 emit_move_insn (reg, const0_rtx);
1178 bytes -= bitsize / BITS_PER_UNIT;
1179 store_bit_field (reg, bitsize, endian_correction, 0, 0,
1180 word_mode, word, false);
1185 /* The limit set by -Walloc-larger-than=. */
1186 static GTY(()) tree alloc_object_size_limit;
1188 /* Initialize ALLOC_OBJECT_SIZE_LIMIT based on the -Walloc-size-larger-than=
1189 setting if the option is specified, or to the maximum object size if it
1190 is not. Return the initialized value. */
1192 static tree
1193 alloc_max_size (void)
1195 if (!alloc_object_size_limit)
1197 alloc_object_size_limit = TYPE_MAX_VALUE (ssizetype);
1199 if (warn_alloc_size_limit)
1201 char *end = NULL;
1202 errno = 0;
1203 unsigned HOST_WIDE_INT unit = 1;
1204 unsigned HOST_WIDE_INT limit
1205 = strtoull (warn_alloc_size_limit, &end, 10);
1207 if (!errno)
1209 if (end && *end)
1211 /* Numeric option arguments are at most INT_MAX. Make it
1212 possible to specify a larger value by accepting common
1213 suffixes. */
1214 if (!strcmp (end, "kB"))
1215 unit = 1000;
1216 else if (!strcasecmp (end, "KiB") || strcmp (end, "KB"))
1217 unit = 1024;
1218 else if (!strcmp (end, "MB"))
1219 unit = HOST_WIDE_INT_UC (1000) * 1000;
1220 else if (!strcasecmp (end, "MiB"))
1221 unit = HOST_WIDE_INT_UC (1024) * 1024;
1222 else if (!strcasecmp (end, "GB"))
1223 unit = HOST_WIDE_INT_UC (1000) * 1000 * 1000;
1224 else if (!strcasecmp (end, "GiB"))
1225 unit = HOST_WIDE_INT_UC (1024) * 1024 * 1024;
1226 else if (!strcasecmp (end, "TB"))
1227 unit = HOST_WIDE_INT_UC (1000) * 1000 * 1000 * 1000;
1228 else if (!strcasecmp (end, "TiB"))
1229 unit = HOST_WIDE_INT_UC (1024) * 1024 * 1024 * 1024;
1230 else if (!strcasecmp (end, "PB"))
1231 unit = HOST_WIDE_INT_UC (1000) * 1000 * 1000 * 1000 * 1000;
1232 else if (!strcasecmp (end, "PiB"))
1233 unit = HOST_WIDE_INT_UC (1024) * 1024 * 1024 * 1024 * 1024;
1234 else if (!strcasecmp (end, "EB"))
1235 unit = HOST_WIDE_INT_UC (1000) * 1000 * 1000 * 1000 * 1000
1236 * 1000;
1237 else if (!strcasecmp (end, "EiB"))
1238 unit = HOST_WIDE_INT_UC (1024) * 1024 * 1024 * 1024 * 1024
1239 * 1024;
1240 else
1241 unit = 0;
1244 if (unit)
1246 widest_int w = wi::mul (limit, unit);
1247 if (w < wi::to_widest (alloc_object_size_limit))
1248 alloc_object_size_limit = wide_int_to_tree (ssizetype, w);
1253 return alloc_object_size_limit;
1256 /* Return true when EXP's range can be determined and set RANGE[] to it
1257 after adjusting it if necessary to make EXP a valid size argument to
1258 an allocation function declared with attribute alloc_size (whose
1259 argument may be signed), or to a string manipulation function like
1260 memset. */
1262 bool
1263 get_size_range (tree exp, tree range[2])
1265 if (tree_fits_uhwi_p (exp))
1267 /* EXP is a constant. */
1268 range[0] = range[1] = exp;
1269 return true;
1272 wide_int min, max;
1273 enum value_range_type range_type
1274 = ((TREE_CODE (exp) == SSA_NAME && INTEGRAL_TYPE_P (TREE_TYPE (exp)))
1275 ? get_range_info (exp, &min, &max) : VR_VARYING);
1277 if (range_type == VR_VARYING)
1279 /* No range information available. */
1280 range[0] = NULL_TREE;
1281 range[1] = NULL_TREE;
1282 return false;
1285 tree exptype = TREE_TYPE (exp);
1286 unsigned expprec = TYPE_PRECISION (exptype);
1288 bool signed_p = !TYPE_UNSIGNED (exptype);
1290 if (range_type == VR_ANTI_RANGE)
1292 if (signed_p)
1294 if (wi::les_p (max, 0))
1296 /* EXP is not in a strictly negative range. That means
1297 it must be in some (not necessarily strictly) positive
1298 range which includes zero. Since in signed to unsigned
1299 conversions negative values end up converted to large
1300 positive values, and otherwise they are not valid sizes,
1301 the resulting range is in both cases [0, TYPE_MAX]. */
1302 min = wi::zero (expprec);
1303 max = wi::to_wide (TYPE_MAX_VALUE (exptype));
1305 else if (wi::les_p (min - 1, 0))
1307 /* EXP is not in a negative-positive range. That means EXP
1308 is either negative, or greater than max. Since negative
1309 sizes are invalid make the range [MAX + 1, TYPE_MAX]. */
1310 min = max + 1;
1311 max = wi::to_wide (TYPE_MAX_VALUE (exptype));
1313 else
1315 max = min - 1;
1316 min = wi::zero (expprec);
1319 else if (wi::eq_p (0, min - 1))
1321 /* EXP is unsigned and not in the range [1, MAX]. That means
1322 it's either zero or greater than MAX. Even though 0 would
1323 normally be detected by -Walloc-zero set the range to
1324 [MAX, TYPE_MAX] so that when MAX is greater than the limit
1325 the whole range is diagnosed. */
1326 min = max + 1;
1327 max = wi::to_wide (TYPE_MAX_VALUE (exptype));
1329 else
1331 max = min - 1;
1332 min = wi::zero (expprec);
1336 range[0] = wide_int_to_tree (exptype, min);
1337 range[1] = wide_int_to_tree (exptype, max);
1339 return true;
1342 /* Diagnose a call EXP to function FN decorated with attribute alloc_size
1343 whose argument numbers given by IDX with values given by ARGS exceed
1344 the maximum object size or cause an unsigned oveflow (wrapping) when
1345 multiplied. When ARGS[0] is null the function does nothing. ARGS[1]
1346 may be null for functions like malloc, and non-null for those like
1347 calloc that are decorated with a two-argument attribute alloc_size. */
1349 void
1350 maybe_warn_alloc_args_overflow (tree fn, tree exp, tree args[2], int idx[2])
1352 /* The range each of the (up to) two arguments is known to be in. */
1353 tree argrange[2][2] = { { NULL_TREE, NULL_TREE }, { NULL_TREE, NULL_TREE } };
1355 /* Maximum object size set by -Walloc-size-larger-than= or SIZE_MAX / 2. */
1356 tree maxobjsize = alloc_max_size ();
1358 location_t loc = EXPR_LOCATION (exp);
1360 bool warned = false;
1362 /* Validate each argument individually. */
1363 for (unsigned i = 0; i != 2 && args[i]; ++i)
1365 if (TREE_CODE (args[i]) == INTEGER_CST)
1367 argrange[i][0] = args[i];
1368 argrange[i][1] = args[i];
1370 if (tree_int_cst_lt (args[i], integer_zero_node))
1372 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1373 "%Kargument %i value %qE is negative",
1374 exp, idx[i] + 1, args[i]);
1376 else if (integer_zerop (args[i]))
1378 /* Avoid issuing -Walloc-zero for allocation functions other
1379 than __builtin_alloca that are declared with attribute
1380 returns_nonnull because there's no portability risk. This
1381 avoids warning for such calls to libiberty's xmalloc and
1382 friends.
1383 Also avoid issuing the warning for calls to function named
1384 "alloca". */
1385 if ((DECL_FUNCTION_CODE (fn) == BUILT_IN_ALLOCA
1386 && IDENTIFIER_LENGTH (DECL_NAME (fn)) != 6)
1387 || (DECL_FUNCTION_CODE (fn) != BUILT_IN_ALLOCA
1388 && !lookup_attribute ("returns_nonnull",
1389 TYPE_ATTRIBUTES (TREE_TYPE (fn)))))
1390 warned = warning_at (loc, OPT_Walloc_zero,
1391 "%Kargument %i value is zero",
1392 exp, idx[i] + 1);
1394 else if (tree_int_cst_lt (maxobjsize, args[i]))
1396 /* G++ emits calls to ::operator new[](SIZE_MAX) in C++98
1397 mode and with -fno-exceptions as a way to indicate array
1398 size overflow. There's no good way to detect C++98 here
1399 so avoid diagnosing these calls for all C++ modes. */
1400 if (i == 0
1401 && !args[1]
1402 && lang_GNU_CXX ()
1403 && DECL_IS_OPERATOR_NEW (fn)
1404 && integer_all_onesp (args[i]))
1405 continue;
1407 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1408 "%Kargument %i value %qE exceeds "
1409 "maximum object size %E",
1410 exp, idx[i] + 1, args[i], maxobjsize);
1413 else if (TREE_CODE (args[i]) == SSA_NAME
1414 && get_size_range (args[i], argrange[i]))
1416 /* Verify that the argument's range is not negative (including
1417 upper bound of zero). */
1418 if (tree_int_cst_lt (argrange[i][0], integer_zero_node)
1419 && tree_int_cst_le (argrange[i][1], integer_zero_node))
1421 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1422 "%Kargument %i range [%E, %E] is negative",
1423 exp, idx[i] + 1,
1424 argrange[i][0], argrange[i][1]);
1426 else if (tree_int_cst_lt (maxobjsize, argrange[i][0]))
1428 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1429 "%Kargument %i range [%E, %E] exceeds "
1430 "maximum object size %E",
1431 exp, idx[i] + 1,
1432 argrange[i][0], argrange[i][1],
1433 maxobjsize);
1438 if (!argrange[0])
1439 return;
1441 /* For a two-argument alloc_size, validate the product of the two
1442 arguments if both of their values or ranges are known. */
1443 if (!warned && tree_fits_uhwi_p (argrange[0][0])
1444 && argrange[1][0] && tree_fits_uhwi_p (argrange[1][0])
1445 && !integer_onep (argrange[0][0])
1446 && !integer_onep (argrange[1][0]))
1448 /* Check for overflow in the product of a function decorated with
1449 attribute alloc_size (X, Y). */
1450 unsigned szprec = TYPE_PRECISION (size_type_node);
1451 wide_int x = wi::to_wide (argrange[0][0], szprec);
1452 wide_int y = wi::to_wide (argrange[1][0], szprec);
1454 bool vflow;
1455 wide_int prod = wi::umul (x, y, &vflow);
1457 if (vflow)
1458 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1459 "%Kproduct %<%E * %E%> of arguments %i and %i "
1460 "exceeds %<SIZE_MAX%>",
1461 exp, argrange[0][0], argrange[1][0],
1462 idx[0] + 1, idx[1] + 1);
1463 else if (wi::ltu_p (wi::to_wide (maxobjsize, szprec), prod))
1464 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1465 "%Kproduct %<%E * %E%> of arguments %i and %i "
1466 "exceeds maximum object size %E",
1467 exp, argrange[0][0], argrange[1][0],
1468 idx[0] + 1, idx[1] + 1,
1469 maxobjsize);
1471 if (warned)
1473 /* Print the full range of each of the two arguments to make
1474 it clear when it is, in fact, in a range and not constant. */
1475 if (argrange[0][0] != argrange [0][1])
1476 inform (loc, "argument %i in the range [%E, %E]",
1477 idx[0] + 1, argrange[0][0], argrange[0][1]);
1478 if (argrange[1][0] != argrange [1][1])
1479 inform (loc, "argument %i in the range [%E, %E]",
1480 idx[1] + 1, argrange[1][0], argrange[1][1]);
1484 if (warned)
1486 location_t fnloc = DECL_SOURCE_LOCATION (fn);
1488 if (DECL_IS_BUILTIN (fn))
1489 inform (loc,
1490 "in a call to built-in allocation function %qD", fn);
1491 else
1492 inform (fnloc,
1493 "in a call to allocation function %qD declared here", fn);
1497 /* Issue an error if CALL_EXPR was flagged as requiring
1498 tall-call optimization. */
1500 static void
1501 maybe_complain_about_tail_call (tree call_expr, const char *reason)
1503 gcc_assert (TREE_CODE (call_expr) == CALL_EXPR);
1504 if (!CALL_EXPR_MUST_TAIL_CALL (call_expr))
1505 return;
1507 error_at (EXPR_LOCATION (call_expr), "cannot tail-call: %s", reason);
1510 /* Fill in ARGS_SIZE and ARGS array based on the parameters found in
1511 CALL_EXPR EXP.
1513 NUM_ACTUALS is the total number of parameters.
1515 N_NAMED_ARGS is the total number of named arguments.
1517 STRUCT_VALUE_ADDR_VALUE is the implicit argument for a struct return
1518 value, or null.
1520 FNDECL is the tree code for the target of this call (if known)
1522 ARGS_SO_FAR holds state needed by the target to know where to place
1523 the next argument.
1525 REG_PARM_STACK_SPACE is the number of bytes of stack space reserved
1526 for arguments which are passed in registers.
1528 OLD_STACK_LEVEL is a pointer to an rtx which olds the old stack level
1529 and may be modified by this routine.
1531 OLD_PENDING_ADJ, MUST_PREALLOCATE and FLAGS are pointers to integer
1532 flags which may be modified by this routine.
1534 MAY_TAILCALL is cleared if we encounter an invisible pass-by-reference
1535 that requires allocation of stack space.
1537 CALL_FROM_THUNK_P is true if this call is the jump from a thunk to
1538 the thunked-to function. */
1540 static void
1541 initialize_argument_information (int num_actuals ATTRIBUTE_UNUSED,
1542 struct arg_data *args,
1543 struct args_size *args_size,
1544 int n_named_args ATTRIBUTE_UNUSED,
1545 tree exp, tree struct_value_addr_value,
1546 tree fndecl, tree fntype,
1547 cumulative_args_t args_so_far,
1548 int reg_parm_stack_space,
1549 rtx *old_stack_level, int *old_pending_adj,
1550 int *must_preallocate, int *ecf_flags,
1551 bool *may_tailcall, bool call_from_thunk_p)
1553 CUMULATIVE_ARGS *args_so_far_pnt = get_cumulative_args (args_so_far);
1554 location_t loc = EXPR_LOCATION (exp);
1556 /* Count arg position in order args appear. */
1557 int argpos;
1559 int i;
1561 args_size->constant = 0;
1562 args_size->var = 0;
1564 bitmap_obstack_initialize (NULL);
1566 /* In this loop, we consider args in the order they are written.
1567 We fill up ARGS from the back. */
1569 i = num_actuals - 1;
1571 int j = i, ptr_arg = -1;
1572 call_expr_arg_iterator iter;
1573 tree arg;
1574 bitmap slots = NULL;
1576 if (struct_value_addr_value)
1578 args[j].tree_value = struct_value_addr_value;
1579 j--;
1581 /* If we pass structure address then we need to
1582 create bounds for it. Since created bounds is
1583 a call statement, we expand it right here to avoid
1584 fixing all other places where it may be expanded. */
1585 if (CALL_WITH_BOUNDS_P (exp))
1587 args[j].value = gen_reg_rtx (targetm.chkp_bound_mode ());
1588 args[j].tree_value
1589 = chkp_make_bounds_for_struct_addr (struct_value_addr_value);
1590 expand_expr_real (args[j].tree_value, args[j].value, VOIDmode,
1591 EXPAND_NORMAL, 0, false);
1592 args[j].pointer_arg = j + 1;
1593 j--;
1596 argpos = 0;
1597 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
1599 tree argtype = TREE_TYPE (arg);
1601 /* Remember last param with pointer and associate it
1602 with following pointer bounds. */
1603 if (CALL_WITH_BOUNDS_P (exp)
1604 && chkp_type_has_pointer (argtype))
1606 if (slots)
1607 BITMAP_FREE (slots);
1608 ptr_arg = j;
1609 if (!BOUNDED_TYPE_P (argtype))
1611 slots = BITMAP_ALLOC (NULL);
1612 chkp_find_bound_slots (argtype, slots);
1615 else if (CALL_WITH_BOUNDS_P (exp)
1616 && pass_by_reference (NULL, TYPE_MODE (argtype), argtype,
1617 argpos < n_named_args))
1619 if (slots)
1620 BITMAP_FREE (slots);
1621 ptr_arg = j;
1623 else if (POINTER_BOUNDS_TYPE_P (argtype))
1625 /* We expect bounds in instrumented calls only.
1626 Otherwise it is a sign we lost flag due to some optimization
1627 and may emit call args incorrectly. */
1628 gcc_assert (CALL_WITH_BOUNDS_P (exp));
1630 /* For structures look for the next available pointer. */
1631 if (ptr_arg != -1 && slots)
1633 unsigned bnd_no = bitmap_first_set_bit (slots);
1634 args[j].pointer_offset =
1635 bnd_no * POINTER_SIZE / BITS_PER_UNIT;
1637 bitmap_clear_bit (slots, bnd_no);
1639 /* Check we have no more pointers in the structure. */
1640 if (bitmap_empty_p (slots))
1641 BITMAP_FREE (slots);
1643 args[j].pointer_arg = ptr_arg;
1645 /* Check we covered all pointers in the previous
1646 non bounds arg. */
1647 if (!slots)
1648 ptr_arg = -1;
1650 else
1651 ptr_arg = -1;
1653 if (targetm.calls.split_complex_arg
1654 && argtype
1655 && TREE_CODE (argtype) == COMPLEX_TYPE
1656 && targetm.calls.split_complex_arg (argtype))
1658 tree subtype = TREE_TYPE (argtype);
1659 args[j].tree_value = build1 (REALPART_EXPR, subtype, arg);
1660 j--;
1661 args[j].tree_value = build1 (IMAGPART_EXPR, subtype, arg);
1663 else
1664 args[j].tree_value = arg;
1665 j--;
1666 argpos++;
1669 if (slots)
1670 BITMAP_FREE (slots);
1673 bitmap_obstack_release (NULL);
1675 /* Extract attribute alloc_size and if set, store the indices of
1676 the corresponding arguments in ALLOC_IDX, and then the actual
1677 argument(s) at those indices in ALLOC_ARGS. */
1678 int alloc_idx[2] = { -1, -1 };
1679 if (tree alloc_size
1680 = (fndecl ? lookup_attribute ("alloc_size",
1681 TYPE_ATTRIBUTES (TREE_TYPE (fndecl)))
1682 : NULL_TREE))
1684 tree args = TREE_VALUE (alloc_size);
1685 alloc_idx[0] = TREE_INT_CST_LOW (TREE_VALUE (args)) - 1;
1686 if (TREE_CHAIN (args))
1687 alloc_idx[1] = TREE_INT_CST_LOW (TREE_VALUE (TREE_CHAIN (args))) - 1;
1690 /* Array for up to the two attribute alloc_size arguments. */
1691 tree alloc_args[] = { NULL_TREE, NULL_TREE };
1693 /* I counts args in order (to be) pushed; ARGPOS counts in order written. */
1694 for (argpos = 0; argpos < num_actuals; i--, argpos++)
1696 tree type = TREE_TYPE (args[i].tree_value);
1697 int unsignedp;
1698 machine_mode mode;
1700 /* Replace erroneous argument with constant zero. */
1701 if (type == error_mark_node || !COMPLETE_TYPE_P (type))
1702 args[i].tree_value = integer_zero_node, type = integer_type_node;
1704 /* If TYPE is a transparent union or record, pass things the way
1705 we would pass the first field of the union or record. We have
1706 already verified that the modes are the same. */
1707 if ((TREE_CODE (type) == UNION_TYPE || TREE_CODE (type) == RECORD_TYPE)
1708 && TYPE_TRANSPARENT_AGGR (type))
1709 type = TREE_TYPE (first_field (type));
1711 /* Decide where to pass this arg.
1713 args[i].reg is nonzero if all or part is passed in registers.
1715 args[i].partial is nonzero if part but not all is passed in registers,
1716 and the exact value says how many bytes are passed in registers.
1718 args[i].pass_on_stack is nonzero if the argument must at least be
1719 computed on the stack. It may then be loaded back into registers
1720 if args[i].reg is nonzero.
1722 These decisions are driven by the FUNCTION_... macros and must agree
1723 with those made by function.c. */
1725 /* See if this argument should be passed by invisible reference. */
1726 if (pass_by_reference (args_so_far_pnt, TYPE_MODE (type),
1727 type, argpos < n_named_args))
1729 bool callee_copies;
1730 tree base = NULL_TREE;
1732 callee_copies
1733 = reference_callee_copied (args_so_far_pnt, TYPE_MODE (type),
1734 type, argpos < n_named_args);
1736 /* If we're compiling a thunk, pass through invisible references
1737 instead of making a copy. */
1738 if (call_from_thunk_p
1739 || (callee_copies
1740 && !TREE_ADDRESSABLE (type)
1741 && (base = get_base_address (args[i].tree_value))
1742 && TREE_CODE (base) != SSA_NAME
1743 && (!DECL_P (base) || MEM_P (DECL_RTL (base)))))
1745 /* We may have turned the parameter value into an SSA name.
1746 Go back to the original parameter so we can take the
1747 address. */
1748 if (TREE_CODE (args[i].tree_value) == SSA_NAME)
1750 gcc_assert (SSA_NAME_IS_DEFAULT_DEF (args[i].tree_value));
1751 args[i].tree_value = SSA_NAME_VAR (args[i].tree_value);
1752 gcc_assert (TREE_CODE (args[i].tree_value) == PARM_DECL);
1754 /* Argument setup code may have copied the value to register. We
1755 revert that optimization now because the tail call code must
1756 use the original location. */
1757 if (TREE_CODE (args[i].tree_value) == PARM_DECL
1758 && !MEM_P (DECL_RTL (args[i].tree_value))
1759 && DECL_INCOMING_RTL (args[i].tree_value)
1760 && MEM_P (DECL_INCOMING_RTL (args[i].tree_value)))
1761 set_decl_rtl (args[i].tree_value,
1762 DECL_INCOMING_RTL (args[i].tree_value));
1764 mark_addressable (args[i].tree_value);
1766 /* We can't use sibcalls if a callee-copied argument is
1767 stored in the current function's frame. */
1768 if (!call_from_thunk_p && DECL_P (base) && !TREE_STATIC (base))
1770 *may_tailcall = false;
1771 maybe_complain_about_tail_call (exp,
1772 "a callee-copied argument is"
1773 " stored in the current "
1774 " function's frame");
1777 args[i].tree_value = build_fold_addr_expr_loc (loc,
1778 args[i].tree_value);
1779 type = TREE_TYPE (args[i].tree_value);
1781 if (*ecf_flags & ECF_CONST)
1782 *ecf_flags &= ~(ECF_CONST | ECF_LOOPING_CONST_OR_PURE);
1784 else
1786 /* We make a copy of the object and pass the address to the
1787 function being called. */
1788 rtx copy;
1790 if (!COMPLETE_TYPE_P (type)
1791 || TREE_CODE (TYPE_SIZE_UNIT (type)) != INTEGER_CST
1792 || (flag_stack_check == GENERIC_STACK_CHECK
1793 && compare_tree_int (TYPE_SIZE_UNIT (type),
1794 STACK_CHECK_MAX_VAR_SIZE) > 0))
1796 /* This is a variable-sized object. Make space on the stack
1797 for it. */
1798 rtx size_rtx = expr_size (args[i].tree_value);
1800 if (*old_stack_level == 0)
1802 emit_stack_save (SAVE_BLOCK, old_stack_level);
1803 *old_pending_adj = pending_stack_adjust;
1804 pending_stack_adjust = 0;
1807 /* We can pass TRUE as the 4th argument because we just
1808 saved the stack pointer and will restore it right after
1809 the call. */
1810 copy = allocate_dynamic_stack_space (size_rtx,
1811 TYPE_ALIGN (type),
1812 TYPE_ALIGN (type),
1813 max_int_size_in_bytes
1814 (type),
1815 true);
1816 copy = gen_rtx_MEM (BLKmode, copy);
1817 set_mem_attributes (copy, type, 1);
1819 else
1820 copy = assign_temp (type, 1, 0);
1822 store_expr (args[i].tree_value, copy, 0, false, false);
1824 /* Just change the const function to pure and then let
1825 the next test clear the pure based on
1826 callee_copies. */
1827 if (*ecf_flags & ECF_CONST)
1829 *ecf_flags &= ~ECF_CONST;
1830 *ecf_flags |= ECF_PURE;
1833 if (!callee_copies && *ecf_flags & ECF_PURE)
1834 *ecf_flags &= ~(ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
1836 args[i].tree_value
1837 = build_fold_addr_expr_loc (loc, make_tree (type, copy));
1838 type = TREE_TYPE (args[i].tree_value);
1839 *may_tailcall = false;
1840 maybe_complain_about_tail_call (exp,
1841 "argument must be passed"
1842 " by copying");
1846 unsignedp = TYPE_UNSIGNED (type);
1847 mode = promote_function_mode (type, TYPE_MODE (type), &unsignedp,
1848 fndecl ? TREE_TYPE (fndecl) : fntype, 0);
1850 args[i].unsignedp = unsignedp;
1851 args[i].mode = mode;
1853 args[i].reg = targetm.calls.function_arg (args_so_far, mode, type,
1854 argpos < n_named_args);
1856 if (args[i].reg && CONST_INT_P (args[i].reg))
1858 args[i].special_slot = args[i].reg;
1859 args[i].reg = NULL;
1862 /* If this is a sibling call and the machine has register windows, the
1863 register window has to be unwinded before calling the routine, so
1864 arguments have to go into the incoming registers. */
1865 if (targetm.calls.function_incoming_arg != targetm.calls.function_arg)
1866 args[i].tail_call_reg
1867 = targetm.calls.function_incoming_arg (args_so_far, mode, type,
1868 argpos < n_named_args);
1869 else
1870 args[i].tail_call_reg = args[i].reg;
1872 if (args[i].reg)
1873 args[i].partial
1874 = targetm.calls.arg_partial_bytes (args_so_far, mode, type,
1875 argpos < n_named_args);
1877 args[i].pass_on_stack = targetm.calls.must_pass_in_stack (mode, type);
1879 /* If FUNCTION_ARG returned a (parallel [(expr_list (nil) ...) ...]),
1880 it means that we are to pass this arg in the register(s) designated
1881 by the PARALLEL, but also to pass it in the stack. */
1882 if (args[i].reg && GET_CODE (args[i].reg) == PARALLEL
1883 && XEXP (XVECEXP (args[i].reg, 0, 0), 0) == 0)
1884 args[i].pass_on_stack = 1;
1886 /* If this is an addressable type, we must preallocate the stack
1887 since we must evaluate the object into its final location.
1889 If this is to be passed in both registers and the stack, it is simpler
1890 to preallocate. */
1891 if (TREE_ADDRESSABLE (type)
1892 || (args[i].pass_on_stack && args[i].reg != 0))
1893 *must_preallocate = 1;
1895 /* No stack allocation and padding for bounds. */
1896 if (POINTER_BOUNDS_P (args[i].tree_value))
1898 /* Compute the stack-size of this argument. */
1899 else if (args[i].reg == 0 || args[i].partial != 0
1900 || reg_parm_stack_space > 0
1901 || args[i].pass_on_stack)
1902 locate_and_pad_parm (mode, type,
1903 #ifdef STACK_PARMS_IN_REG_PARM_AREA
1905 #else
1906 args[i].reg != 0,
1907 #endif
1908 reg_parm_stack_space,
1909 args[i].pass_on_stack ? 0 : args[i].partial,
1910 fndecl, args_size, &args[i].locate);
1911 #ifdef BLOCK_REG_PADDING
1912 else
1913 /* The argument is passed entirely in registers. See at which
1914 end it should be padded. */
1915 args[i].locate.where_pad =
1916 BLOCK_REG_PADDING (mode, type,
1917 int_size_in_bytes (type) <= UNITS_PER_WORD);
1918 #endif
1920 /* Update ARGS_SIZE, the total stack space for args so far. */
1922 args_size->constant += args[i].locate.size.constant;
1923 if (args[i].locate.size.var)
1924 ADD_PARM_SIZE (*args_size, args[i].locate.size.var);
1926 /* Increment ARGS_SO_FAR, which has info about which arg-registers
1927 have been used, etc. */
1929 targetm.calls.function_arg_advance (args_so_far, TYPE_MODE (type),
1930 type, argpos < n_named_args);
1932 /* Store argument values for functions decorated with attribute
1933 alloc_size. */
1934 if (argpos == alloc_idx[0])
1935 alloc_args[0] = args[i].tree_value;
1936 else if (argpos == alloc_idx[1])
1937 alloc_args[1] = args[i].tree_value;
1940 if (alloc_args[0])
1942 /* Check the arguments of functions decorated with attribute
1943 alloc_size. */
1944 maybe_warn_alloc_args_overflow (fndecl, exp, alloc_args, alloc_idx);
1948 /* Update ARGS_SIZE to contain the total size for the argument block.
1949 Return the original constant component of the argument block's size.
1951 REG_PARM_STACK_SPACE holds the number of bytes of stack space reserved
1952 for arguments passed in registers. */
1954 static int
1955 compute_argument_block_size (int reg_parm_stack_space,
1956 struct args_size *args_size,
1957 tree fndecl ATTRIBUTE_UNUSED,
1958 tree fntype ATTRIBUTE_UNUSED,
1959 int preferred_stack_boundary ATTRIBUTE_UNUSED)
1961 int unadjusted_args_size = args_size->constant;
1963 /* For accumulate outgoing args mode we don't need to align, since the frame
1964 will be already aligned. Align to STACK_BOUNDARY in order to prevent
1965 backends from generating misaligned frame sizes. */
1966 if (ACCUMULATE_OUTGOING_ARGS && preferred_stack_boundary > STACK_BOUNDARY)
1967 preferred_stack_boundary = STACK_BOUNDARY;
1969 /* Compute the actual size of the argument block required. The variable
1970 and constant sizes must be combined, the size may have to be rounded,
1971 and there may be a minimum required size. */
1973 if (args_size->var)
1975 args_size->var = ARGS_SIZE_TREE (*args_size);
1976 args_size->constant = 0;
1978 preferred_stack_boundary /= BITS_PER_UNIT;
1979 if (preferred_stack_boundary > 1)
1981 /* We don't handle this case yet. To handle it correctly we have
1982 to add the delta, round and subtract the delta.
1983 Currently no machine description requires this support. */
1984 gcc_assert (!(stack_pointer_delta & (preferred_stack_boundary - 1)));
1985 args_size->var = round_up (args_size->var, preferred_stack_boundary);
1988 if (reg_parm_stack_space > 0)
1990 args_size->var
1991 = size_binop (MAX_EXPR, args_size->var,
1992 ssize_int (reg_parm_stack_space));
1994 /* The area corresponding to register parameters is not to count in
1995 the size of the block we need. So make the adjustment. */
1996 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
1997 args_size->var
1998 = size_binop (MINUS_EXPR, args_size->var,
1999 ssize_int (reg_parm_stack_space));
2002 else
2004 preferred_stack_boundary /= BITS_PER_UNIT;
2005 if (preferred_stack_boundary < 1)
2006 preferred_stack_boundary = 1;
2007 args_size->constant = (((args_size->constant
2008 + stack_pointer_delta
2009 + preferred_stack_boundary - 1)
2010 / preferred_stack_boundary
2011 * preferred_stack_boundary)
2012 - stack_pointer_delta);
2014 args_size->constant = MAX (args_size->constant,
2015 reg_parm_stack_space);
2017 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
2018 args_size->constant -= reg_parm_stack_space;
2020 return unadjusted_args_size;
2023 /* Precompute parameters as needed for a function call.
2025 FLAGS is mask of ECF_* constants.
2027 NUM_ACTUALS is the number of arguments.
2029 ARGS is an array containing information for each argument; this
2030 routine fills in the INITIAL_VALUE and VALUE fields for each
2031 precomputed argument. */
2033 static void
2034 precompute_arguments (int num_actuals, struct arg_data *args)
2036 int i;
2038 /* If this is a libcall, then precompute all arguments so that we do not
2039 get extraneous instructions emitted as part of the libcall sequence. */
2041 /* If we preallocated the stack space, and some arguments must be passed
2042 on the stack, then we must precompute any parameter which contains a
2043 function call which will store arguments on the stack.
2044 Otherwise, evaluating the parameter may clobber previous parameters
2045 which have already been stored into the stack. (we have code to avoid
2046 such case by saving the outgoing stack arguments, but it results in
2047 worse code) */
2048 if (!ACCUMULATE_OUTGOING_ARGS)
2049 return;
2051 for (i = 0; i < num_actuals; i++)
2053 tree type;
2054 machine_mode mode;
2056 if (TREE_CODE (args[i].tree_value) != CALL_EXPR)
2057 continue;
2059 /* If this is an addressable type, we cannot pre-evaluate it. */
2060 type = TREE_TYPE (args[i].tree_value);
2061 gcc_assert (!TREE_ADDRESSABLE (type));
2063 args[i].initial_value = args[i].value
2064 = expand_normal (args[i].tree_value);
2066 mode = TYPE_MODE (type);
2067 if (mode != args[i].mode)
2069 int unsignedp = args[i].unsignedp;
2070 args[i].value
2071 = convert_modes (args[i].mode, mode,
2072 args[i].value, args[i].unsignedp);
2074 /* CSE will replace this only if it contains args[i].value
2075 pseudo, so convert it down to the declared mode using
2076 a SUBREG. */
2077 if (REG_P (args[i].value)
2078 && GET_MODE_CLASS (args[i].mode) == MODE_INT
2079 && promote_mode (type, mode, &unsignedp) != args[i].mode)
2081 args[i].initial_value
2082 = gen_lowpart_SUBREG (mode, args[i].value);
2083 SUBREG_PROMOTED_VAR_P (args[i].initial_value) = 1;
2084 SUBREG_PROMOTED_SET (args[i].initial_value, args[i].unsignedp);
2090 /* Given the current state of MUST_PREALLOCATE and information about
2091 arguments to a function call in NUM_ACTUALS, ARGS and ARGS_SIZE,
2092 compute and return the final value for MUST_PREALLOCATE. */
2094 static int
2095 finalize_must_preallocate (int must_preallocate, int num_actuals,
2096 struct arg_data *args, struct args_size *args_size)
2098 /* See if we have or want to preallocate stack space.
2100 If we would have to push a partially-in-regs parm
2101 before other stack parms, preallocate stack space instead.
2103 If the size of some parm is not a multiple of the required stack
2104 alignment, we must preallocate.
2106 If the total size of arguments that would otherwise create a copy in
2107 a temporary (such as a CALL) is more than half the total argument list
2108 size, preallocation is faster.
2110 Another reason to preallocate is if we have a machine (like the m88k)
2111 where stack alignment is required to be maintained between every
2112 pair of insns, not just when the call is made. However, we assume here
2113 that such machines either do not have push insns (and hence preallocation
2114 would occur anyway) or the problem is taken care of with
2115 PUSH_ROUNDING. */
2117 if (! must_preallocate)
2119 int partial_seen = 0;
2120 int copy_to_evaluate_size = 0;
2121 int i;
2123 for (i = 0; i < num_actuals && ! must_preallocate; i++)
2125 if (args[i].partial > 0 && ! args[i].pass_on_stack)
2126 partial_seen = 1;
2127 else if (partial_seen && args[i].reg == 0)
2128 must_preallocate = 1;
2129 /* We preallocate in case there are bounds passed
2130 in the bounds table to have precomputed address
2131 for bounds association. */
2132 else if (POINTER_BOUNDS_P (args[i].tree_value)
2133 && !args[i].reg)
2134 must_preallocate = 1;
2136 if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode
2137 && (TREE_CODE (args[i].tree_value) == CALL_EXPR
2138 || TREE_CODE (args[i].tree_value) == TARGET_EXPR
2139 || TREE_CODE (args[i].tree_value) == COND_EXPR
2140 || TREE_ADDRESSABLE (TREE_TYPE (args[i].tree_value))))
2141 copy_to_evaluate_size
2142 += int_size_in_bytes (TREE_TYPE (args[i].tree_value));
2145 if (copy_to_evaluate_size * 2 >= args_size->constant
2146 && args_size->constant > 0)
2147 must_preallocate = 1;
2149 return must_preallocate;
2152 /* If we preallocated stack space, compute the address of each argument
2153 and store it into the ARGS array.
2155 We need not ensure it is a valid memory address here; it will be
2156 validized when it is used.
2158 ARGBLOCK is an rtx for the address of the outgoing arguments. */
2160 static void
2161 compute_argument_addresses (struct arg_data *args, rtx argblock, int num_actuals)
2163 if (argblock)
2165 rtx arg_reg = argblock;
2166 int i, arg_offset = 0;
2168 if (GET_CODE (argblock) == PLUS)
2169 arg_reg = XEXP (argblock, 0), arg_offset = INTVAL (XEXP (argblock, 1));
2171 for (i = 0; i < num_actuals; i++)
2173 rtx offset = ARGS_SIZE_RTX (args[i].locate.offset);
2174 rtx slot_offset = ARGS_SIZE_RTX (args[i].locate.slot_offset);
2175 rtx addr;
2176 unsigned int align, boundary;
2177 unsigned int units_on_stack = 0;
2178 machine_mode partial_mode = VOIDmode;
2180 /* Skip this parm if it will not be passed on the stack. */
2181 if (! args[i].pass_on_stack
2182 && args[i].reg != 0
2183 && args[i].partial == 0)
2184 continue;
2186 /* Pointer Bounds are never passed on the stack. */
2187 if (POINTER_BOUNDS_P (args[i].tree_value))
2188 continue;
2190 addr = simplify_gen_binary (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 unsigned int bits_on_stack = units_on_stack * BITS_PER_UNIT;
2199 partial_mode = int_mode_for_size (bits_on_stack, 1).else_blk ();
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 != 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 addr = simplify_gen_binary (PLUS, Pmode, arg_reg, slot_offset);
2221 addr = plus_constant (Pmode, addr, arg_offset);
2223 if (args[i].partial != 0)
2225 /* Only part of the parameter is being passed on the stack.
2226 Generate a simple memory reference of the correct size.
2228 args[i].stack_slot = gen_rtx_MEM (partial_mode, addr);
2229 set_mem_size (args[i].stack_slot, units_on_stack);
2231 else
2233 args[i].stack_slot = gen_rtx_MEM (args[i].mode, addr);
2234 set_mem_attributes (args[i].stack_slot,
2235 TREE_TYPE (args[i].tree_value), 1);
2237 set_mem_align (args[i].stack_slot, args[i].locate.boundary);
2239 /* Function incoming arguments may overlap with sibling call
2240 outgoing arguments and we cannot allow reordering of reads
2241 from function arguments with stores to outgoing arguments
2242 of sibling calls. */
2243 set_mem_alias_set (args[i].stack, 0);
2244 set_mem_alias_set (args[i].stack_slot, 0);
2249 /* Given a FNDECL and EXP, return an rtx suitable for use as a target address
2250 in a call instruction.
2252 FNDECL is the tree node for the target function. For an indirect call
2253 FNDECL will be NULL_TREE.
2255 ADDR is the operand 0 of CALL_EXPR for this call. */
2257 static rtx
2258 rtx_for_function_call (tree fndecl, tree addr)
2260 rtx funexp;
2262 /* Get the function to call, in the form of RTL. */
2263 if (fndecl)
2265 if (!TREE_USED (fndecl) && fndecl != current_function_decl)
2266 TREE_USED (fndecl) = 1;
2268 /* Get a SYMBOL_REF rtx for the function address. */
2269 funexp = XEXP (DECL_RTL (fndecl), 0);
2271 else
2272 /* Generate an rtx (probably a pseudo-register) for the address. */
2274 push_temp_slots ();
2275 funexp = expand_normal (addr);
2276 pop_temp_slots (); /* FUNEXP can't be BLKmode. */
2278 return funexp;
2281 /* Internal state for internal_arg_pointer_based_exp and its helpers. */
2282 static struct
2284 /* Last insn that has been scanned by internal_arg_pointer_based_exp_scan,
2285 or NULL_RTX if none has been scanned yet. */
2286 rtx_insn *scan_start;
2287 /* Vector indexed by REGNO - FIRST_PSEUDO_REGISTER, recording if a pseudo is
2288 based on crtl->args.internal_arg_pointer. The element is NULL_RTX if the
2289 pseudo isn't based on it, a CONST_INT offset if the pseudo is based on it
2290 with fixed offset, or PC if this is with variable or unknown offset. */
2291 vec<rtx> cache;
2292 } internal_arg_pointer_exp_state;
2294 static rtx internal_arg_pointer_based_exp (const_rtx, bool);
2296 /* Helper function for internal_arg_pointer_based_exp. Scan insns in
2297 the tail call sequence, starting with first insn that hasn't been
2298 scanned yet, and note for each pseudo on the LHS whether it is based
2299 on crtl->args.internal_arg_pointer or not, and what offset from that
2300 that pointer it has. */
2302 static void
2303 internal_arg_pointer_based_exp_scan (void)
2305 rtx_insn *insn, *scan_start = internal_arg_pointer_exp_state.scan_start;
2307 if (scan_start == NULL_RTX)
2308 insn = get_insns ();
2309 else
2310 insn = NEXT_INSN (scan_start);
2312 while (insn)
2314 rtx set = single_set (insn);
2315 if (set && REG_P (SET_DEST (set)) && !HARD_REGISTER_P (SET_DEST (set)))
2317 rtx val = NULL_RTX;
2318 unsigned int idx = REGNO (SET_DEST (set)) - FIRST_PSEUDO_REGISTER;
2319 /* Punt on pseudos set multiple times. */
2320 if (idx < internal_arg_pointer_exp_state.cache.length ()
2321 && (internal_arg_pointer_exp_state.cache[idx]
2322 != NULL_RTX))
2323 val = pc_rtx;
2324 else
2325 val = internal_arg_pointer_based_exp (SET_SRC (set), false);
2326 if (val != NULL_RTX)
2328 if (idx >= internal_arg_pointer_exp_state.cache.length ())
2329 internal_arg_pointer_exp_state.cache
2330 .safe_grow_cleared (idx + 1);
2331 internal_arg_pointer_exp_state.cache[idx] = val;
2334 if (NEXT_INSN (insn) == NULL_RTX)
2335 scan_start = insn;
2336 insn = NEXT_INSN (insn);
2339 internal_arg_pointer_exp_state.scan_start = scan_start;
2342 /* Compute whether RTL is based on crtl->args.internal_arg_pointer. Return
2343 NULL_RTX if RTL isn't based on it, a CONST_INT offset if RTL is based on
2344 it with fixed offset, or PC if this is with variable or unknown offset.
2345 TOPLEVEL is true if the function is invoked at the topmost level. */
2347 static rtx
2348 internal_arg_pointer_based_exp (const_rtx rtl, bool toplevel)
2350 if (CONSTANT_P (rtl))
2351 return NULL_RTX;
2353 if (rtl == crtl->args.internal_arg_pointer)
2354 return const0_rtx;
2356 if (REG_P (rtl) && HARD_REGISTER_P (rtl))
2357 return NULL_RTX;
2359 if (GET_CODE (rtl) == PLUS && CONST_INT_P (XEXP (rtl, 1)))
2361 rtx val = internal_arg_pointer_based_exp (XEXP (rtl, 0), toplevel);
2362 if (val == NULL_RTX || val == pc_rtx)
2363 return val;
2364 return plus_constant (Pmode, val, INTVAL (XEXP (rtl, 1)));
2367 /* When called at the topmost level, scan pseudo assignments in between the
2368 last scanned instruction in the tail call sequence and the latest insn
2369 in that sequence. */
2370 if (toplevel)
2371 internal_arg_pointer_based_exp_scan ();
2373 if (REG_P (rtl))
2375 unsigned int idx = REGNO (rtl) - FIRST_PSEUDO_REGISTER;
2376 if (idx < internal_arg_pointer_exp_state.cache.length ())
2377 return internal_arg_pointer_exp_state.cache[idx];
2379 return NULL_RTX;
2382 subrtx_iterator::array_type array;
2383 FOR_EACH_SUBRTX (iter, array, rtl, NONCONST)
2385 const_rtx x = *iter;
2386 if (REG_P (x) && internal_arg_pointer_based_exp (x, false) != NULL_RTX)
2387 return pc_rtx;
2388 if (MEM_P (x))
2389 iter.skip_subrtxes ();
2392 return NULL_RTX;
2395 /* Return true if and only if SIZE storage units (usually bytes)
2396 starting from address ADDR overlap with already clobbered argument
2397 area. This function is used to determine if we should give up a
2398 sibcall. */
2400 static bool
2401 mem_overlaps_already_clobbered_arg_p (rtx addr, unsigned HOST_WIDE_INT size)
2403 HOST_WIDE_INT i;
2404 rtx val;
2406 if (bitmap_empty_p (stored_args_map))
2407 return false;
2408 val = internal_arg_pointer_based_exp (addr, true);
2409 if (val == NULL_RTX)
2410 return false;
2411 else if (val == pc_rtx)
2412 return true;
2413 else
2414 i = INTVAL (val);
2416 if (STACK_GROWS_DOWNWARD)
2417 i -= crtl->args.pretend_args_size;
2418 else
2419 i += crtl->args.pretend_args_size;
2422 if (ARGS_GROW_DOWNWARD)
2423 i = -i - size;
2425 if (size > 0)
2427 unsigned HOST_WIDE_INT k;
2429 for (k = 0; k < size; k++)
2430 if (i + k < SBITMAP_SIZE (stored_args_map)
2431 && bitmap_bit_p (stored_args_map, i + k))
2432 return true;
2435 return false;
2438 /* Do the register loads required for any wholly-register parms or any
2439 parms which are passed both on the stack and in a register. Their
2440 expressions were already evaluated.
2442 Mark all register-parms as living through the call, putting these USE
2443 insns in the CALL_INSN_FUNCTION_USAGE field.
2445 When IS_SIBCALL, perform the check_sibcall_argument_overlap
2446 checking, setting *SIBCALL_FAILURE if appropriate. */
2448 static void
2449 load_register_parameters (struct arg_data *args, int num_actuals,
2450 rtx *call_fusage, int flags, int is_sibcall,
2451 int *sibcall_failure)
2453 int i, j;
2455 for (i = 0; i < num_actuals; i++)
2457 rtx reg = ((flags & ECF_SIBCALL)
2458 ? args[i].tail_call_reg : args[i].reg);
2459 if (reg)
2461 int partial = args[i].partial;
2462 int nregs;
2463 int size = 0;
2464 rtx_insn *before_arg = get_last_insn ();
2465 /* Set non-negative if we must move a word at a time, even if
2466 just one word (e.g, partial == 4 && mode == DFmode). Set
2467 to -1 if we just use a normal move insn. This value can be
2468 zero if the argument is a zero size structure. */
2469 nregs = -1;
2470 if (GET_CODE (reg) == PARALLEL)
2472 else if (partial)
2474 gcc_assert (partial % UNITS_PER_WORD == 0);
2475 nregs = partial / UNITS_PER_WORD;
2477 else if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode)
2479 size = int_size_in_bytes (TREE_TYPE (args[i].tree_value));
2480 nregs = (size + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD;
2482 else
2483 size = GET_MODE_SIZE (args[i].mode);
2485 /* Handle calls that pass values in multiple non-contiguous
2486 locations. The Irix 6 ABI has examples of this. */
2488 if (GET_CODE (reg) == PARALLEL)
2489 emit_group_move (reg, args[i].parallel_value);
2491 /* If simple case, just do move. If normal partial, store_one_arg
2492 has already loaded the register for us. In all other cases,
2493 load the register(s) from memory. */
2495 else if (nregs == -1)
2497 emit_move_insn (reg, args[i].value);
2498 #ifdef BLOCK_REG_PADDING
2499 /* Handle case where we have a value that needs shifting
2500 up to the msb. eg. a QImode value and we're padding
2501 upward on a BYTES_BIG_ENDIAN machine. */
2502 if (size < UNITS_PER_WORD
2503 && (args[i].locate.where_pad
2504 == (BYTES_BIG_ENDIAN ? PAD_UPWARD : PAD_DOWNWARD)))
2506 rtx x;
2507 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
2509 /* Assigning REG here rather than a temp makes CALL_FUSAGE
2510 report the whole reg as used. Strictly speaking, the
2511 call only uses SIZE bytes at the msb end, but it doesn't
2512 seem worth generating rtl to say that. */
2513 reg = gen_rtx_REG (word_mode, REGNO (reg));
2514 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1);
2515 if (x != reg)
2516 emit_move_insn (reg, x);
2518 #endif
2521 /* If we have pre-computed the values to put in the registers in
2522 the case of non-aligned structures, copy them in now. */
2524 else if (args[i].n_aligned_regs != 0)
2525 for (j = 0; j < args[i].n_aligned_regs; j++)
2526 emit_move_insn (gen_rtx_REG (word_mode, REGNO (reg) + j),
2527 args[i].aligned_regs[j]);
2529 else if (partial == 0 || args[i].pass_on_stack)
2531 rtx mem = validize_mem (copy_rtx (args[i].value));
2533 /* Check for overlap with already clobbered argument area,
2534 providing that this has non-zero size. */
2535 if (is_sibcall
2536 && size != 0
2537 && (mem_overlaps_already_clobbered_arg_p
2538 (XEXP (args[i].value, 0), size)))
2539 *sibcall_failure = 1;
2541 if (size % UNITS_PER_WORD == 0
2542 || MEM_ALIGN (mem) % BITS_PER_WORD == 0)
2543 move_block_to_reg (REGNO (reg), mem, nregs, args[i].mode);
2544 else
2546 if (nregs > 1)
2547 move_block_to_reg (REGNO (reg), mem, nregs - 1,
2548 args[i].mode);
2549 rtx dest = gen_rtx_REG (word_mode, REGNO (reg) + nregs - 1);
2550 unsigned int bitoff = (nregs - 1) * BITS_PER_WORD;
2551 unsigned int bitsize = size * BITS_PER_UNIT - bitoff;
2552 rtx x = extract_bit_field (mem, bitsize, bitoff, 1, dest,
2553 word_mode, word_mode, false,
2554 NULL);
2555 if (BYTES_BIG_ENDIAN)
2556 x = expand_shift (LSHIFT_EXPR, word_mode, x,
2557 BITS_PER_WORD - bitsize, dest, 1);
2558 if (x != dest)
2559 emit_move_insn (dest, x);
2562 /* Handle a BLKmode that needs shifting. */
2563 if (nregs == 1 && size < UNITS_PER_WORD
2564 #ifdef BLOCK_REG_PADDING
2565 && args[i].locate.where_pad == PAD_DOWNWARD
2566 #else
2567 && BYTES_BIG_ENDIAN
2568 #endif
2571 rtx dest = gen_rtx_REG (word_mode, REGNO (reg));
2572 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
2573 enum tree_code dir = (BYTES_BIG_ENDIAN
2574 ? RSHIFT_EXPR : LSHIFT_EXPR);
2575 rtx x;
2577 x = expand_shift (dir, word_mode, dest, shift, dest, 1);
2578 if (x != dest)
2579 emit_move_insn (dest, x);
2583 /* When a parameter is a block, and perhaps in other cases, it is
2584 possible that it did a load from an argument slot that was
2585 already clobbered. */
2586 if (is_sibcall
2587 && check_sibcall_argument_overlap (before_arg, &args[i], 0))
2588 *sibcall_failure = 1;
2590 /* Handle calls that pass values in multiple non-contiguous
2591 locations. The Irix 6 ABI has examples of this. */
2592 if (GET_CODE (reg) == PARALLEL)
2593 use_group_regs (call_fusage, reg);
2594 else if (nregs == -1)
2595 use_reg_mode (call_fusage, reg,
2596 TYPE_MODE (TREE_TYPE (args[i].tree_value)));
2597 else if (nregs > 0)
2598 use_regs (call_fusage, REGNO (reg), nregs);
2603 /* We need to pop PENDING_STACK_ADJUST bytes. But, if the arguments
2604 wouldn't fill up an even multiple of PREFERRED_UNIT_STACK_BOUNDARY
2605 bytes, then we would need to push some additional bytes to pad the
2606 arguments. So, we compute an adjust to the stack pointer for an
2607 amount that will leave the stack under-aligned by UNADJUSTED_ARGS_SIZE
2608 bytes. Then, when the arguments are pushed the stack will be perfectly
2609 aligned. ARGS_SIZE->CONSTANT is set to the number of bytes that should
2610 be popped after the call. Returns the adjustment. */
2612 static int
2613 combine_pending_stack_adjustment_and_call (int unadjusted_args_size,
2614 struct args_size *args_size,
2615 unsigned int preferred_unit_stack_boundary)
2617 /* The number of bytes to pop so that the stack will be
2618 under-aligned by UNADJUSTED_ARGS_SIZE bytes. */
2619 HOST_WIDE_INT adjustment;
2620 /* The alignment of the stack after the arguments are pushed, if we
2621 just pushed the arguments without adjust the stack here. */
2622 unsigned HOST_WIDE_INT unadjusted_alignment;
2624 unadjusted_alignment
2625 = ((stack_pointer_delta + unadjusted_args_size)
2626 % preferred_unit_stack_boundary);
2628 /* We want to get rid of as many of the PENDING_STACK_ADJUST bytes
2629 as possible -- leaving just enough left to cancel out the
2630 UNADJUSTED_ALIGNMENT. In other words, we want to ensure that the
2631 PENDING_STACK_ADJUST is non-negative, and congruent to
2632 -UNADJUSTED_ALIGNMENT modulo the PREFERRED_UNIT_STACK_BOUNDARY. */
2634 /* Begin by trying to pop all the bytes. */
2635 unadjusted_alignment
2636 = (unadjusted_alignment
2637 - (pending_stack_adjust % preferred_unit_stack_boundary));
2638 adjustment = pending_stack_adjust;
2639 /* Push enough additional bytes that the stack will be aligned
2640 after the arguments are pushed. */
2641 if (preferred_unit_stack_boundary > 1 && unadjusted_alignment)
2642 adjustment -= preferred_unit_stack_boundary - unadjusted_alignment;
2644 /* Now, sets ARGS_SIZE->CONSTANT so that we pop the right number of
2645 bytes after the call. The right number is the entire
2646 PENDING_STACK_ADJUST less our ADJUSTMENT plus the amount required
2647 by the arguments in the first place. */
2648 args_size->constant
2649 = pending_stack_adjust - adjustment + unadjusted_args_size;
2651 return adjustment;
2654 /* Scan X expression if it does not dereference any argument slots
2655 we already clobbered by tail call arguments (as noted in stored_args_map
2656 bitmap).
2657 Return nonzero if X expression dereferences such argument slots,
2658 zero otherwise. */
2660 static int
2661 check_sibcall_argument_overlap_1 (rtx x)
2663 RTX_CODE code;
2664 int i, j;
2665 const char *fmt;
2667 if (x == NULL_RTX)
2668 return 0;
2670 code = GET_CODE (x);
2672 /* We need not check the operands of the CALL expression itself. */
2673 if (code == CALL)
2674 return 0;
2676 if (code == MEM)
2677 return mem_overlaps_already_clobbered_arg_p (XEXP (x, 0),
2678 GET_MODE_SIZE (GET_MODE (x)));
2680 /* Scan all subexpressions. */
2681 fmt = GET_RTX_FORMAT (code);
2682 for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++)
2684 if (*fmt == 'e')
2686 if (check_sibcall_argument_overlap_1 (XEXP (x, i)))
2687 return 1;
2689 else if (*fmt == 'E')
2691 for (j = 0; j < XVECLEN (x, i); j++)
2692 if (check_sibcall_argument_overlap_1 (XVECEXP (x, i, j)))
2693 return 1;
2696 return 0;
2699 /* Scan sequence after INSN if it does not dereference any argument slots
2700 we already clobbered by tail call arguments (as noted in stored_args_map
2701 bitmap). If MARK_STORED_ARGS_MAP, add stack slots for ARG to
2702 stored_args_map bitmap afterwards (when ARG is a register MARK_STORED_ARGS_MAP
2703 should be 0). Return nonzero if sequence after INSN dereferences such argument
2704 slots, zero otherwise. */
2706 static int
2707 check_sibcall_argument_overlap (rtx_insn *insn, struct arg_data *arg,
2708 int mark_stored_args_map)
2710 int low, high;
2712 if (insn == NULL_RTX)
2713 insn = get_insns ();
2714 else
2715 insn = NEXT_INSN (insn);
2717 for (; insn; insn = NEXT_INSN (insn))
2718 if (INSN_P (insn)
2719 && check_sibcall_argument_overlap_1 (PATTERN (insn)))
2720 break;
2722 if (mark_stored_args_map)
2724 if (ARGS_GROW_DOWNWARD)
2725 low = -arg->locate.slot_offset.constant - arg->locate.size.constant;
2726 else
2727 low = arg->locate.slot_offset.constant;
2729 for (high = low + arg->locate.size.constant; low < high; low++)
2730 bitmap_set_bit (stored_args_map, low);
2732 return insn != NULL_RTX;
2735 /* Given that a function returns a value of mode MODE at the most
2736 significant end of hard register VALUE, shift VALUE left or right
2737 as specified by LEFT_P. Return true if some action was needed. */
2739 bool
2740 shift_return_value (machine_mode mode, bool left_p, rtx value)
2742 HOST_WIDE_INT shift;
2744 gcc_assert (REG_P (value) && HARD_REGISTER_P (value));
2745 shift = GET_MODE_BITSIZE (GET_MODE (value)) - GET_MODE_BITSIZE (mode);
2746 if (shift == 0)
2747 return false;
2749 /* Use ashr rather than lshr for right shifts. This is for the benefit
2750 of the MIPS port, which requires SImode values to be sign-extended
2751 when stored in 64-bit registers. */
2752 if (!force_expand_binop (GET_MODE (value), left_p ? ashl_optab : ashr_optab,
2753 value, GEN_INT (shift), value, 1, OPTAB_WIDEN))
2754 gcc_unreachable ();
2755 return true;
2758 /* If X is a likely-spilled register value, copy it to a pseudo
2759 register and return that register. Return X otherwise. */
2761 static rtx
2762 avoid_likely_spilled_reg (rtx x)
2764 rtx new_rtx;
2766 if (REG_P (x)
2767 && HARD_REGISTER_P (x)
2768 && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (x))))
2770 /* Make sure that we generate a REG rather than a CONCAT.
2771 Moves into CONCATs can need nontrivial instructions,
2772 and the whole point of this function is to avoid
2773 using the hard register directly in such a situation. */
2774 generating_concat_p = 0;
2775 new_rtx = gen_reg_rtx (GET_MODE (x));
2776 generating_concat_p = 1;
2777 emit_move_insn (new_rtx, x);
2778 return new_rtx;
2780 return x;
2783 /* Helper function for expand_call.
2784 Return false is EXP is not implementable as a sibling call. */
2786 static bool
2787 can_implement_as_sibling_call_p (tree exp,
2788 rtx structure_value_addr,
2789 tree funtype,
2790 int reg_parm_stack_space ATTRIBUTE_UNUSED,
2791 tree fndecl,
2792 int flags,
2793 tree addr,
2794 const args_size &args_size)
2796 if (!targetm.have_sibcall_epilogue ())
2798 maybe_complain_about_tail_call
2799 (exp,
2800 "machine description does not have"
2801 " a sibcall_epilogue instruction pattern");
2802 return false;
2805 /* Doing sibling call optimization needs some work, since
2806 structure_value_addr can be allocated on the stack.
2807 It does not seem worth the effort since few optimizable
2808 sibling calls will return a structure. */
2809 if (structure_value_addr != NULL_RTX)
2811 maybe_complain_about_tail_call (exp, "callee returns a structure");
2812 return false;
2815 #ifdef REG_PARM_STACK_SPACE
2816 /* If outgoing reg parm stack space changes, we can not do sibcall. */
2817 if (OUTGOING_REG_PARM_STACK_SPACE (funtype)
2818 != OUTGOING_REG_PARM_STACK_SPACE (TREE_TYPE (current_function_decl))
2819 || (reg_parm_stack_space != REG_PARM_STACK_SPACE (current_function_decl)))
2821 maybe_complain_about_tail_call (exp,
2822 "inconsistent size of stack space"
2823 " allocated for arguments which are"
2824 " passed in registers");
2825 return false;
2827 #endif
2829 /* Check whether the target is able to optimize the call
2830 into a sibcall. */
2831 if (!targetm.function_ok_for_sibcall (fndecl, exp))
2833 maybe_complain_about_tail_call (exp,
2834 "target is not able to optimize the"
2835 " call into a sibling call");
2836 return false;
2839 /* Functions that do not return exactly once may not be sibcall
2840 optimized. */
2841 if (flags & ECF_RETURNS_TWICE)
2843 maybe_complain_about_tail_call (exp, "callee returns twice");
2844 return false;
2846 if (flags & ECF_NORETURN)
2848 maybe_complain_about_tail_call (exp, "callee does not return");
2849 return false;
2852 if (TYPE_VOLATILE (TREE_TYPE (TREE_TYPE (addr))))
2854 maybe_complain_about_tail_call (exp, "volatile function type");
2855 return false;
2858 /* If the called function is nested in the current one, it might access
2859 some of the caller's arguments, but could clobber them beforehand if
2860 the argument areas are shared. */
2861 if (fndecl && decl_function_context (fndecl) == current_function_decl)
2863 maybe_complain_about_tail_call (exp, "nested function");
2864 return false;
2867 /* If this function requires more stack slots than the current
2868 function, we cannot change it into a sibling call.
2869 crtl->args.pretend_args_size is not part of the
2870 stack allocated by our caller. */
2871 if (args_size.constant > (crtl->args.size - crtl->args.pretend_args_size))
2873 maybe_complain_about_tail_call (exp,
2874 "callee required more stack slots"
2875 " than the caller");
2876 return false;
2879 /* If the callee pops its own arguments, then it must pop exactly
2880 the same number of arguments as the current function. */
2881 if (targetm.calls.return_pops_args (fndecl, funtype, args_size.constant)
2882 != targetm.calls.return_pops_args (current_function_decl,
2883 TREE_TYPE (current_function_decl),
2884 crtl->args.size))
2886 maybe_complain_about_tail_call (exp,
2887 "inconsistent number of"
2888 " popped arguments");
2889 return false;
2892 if (!lang_hooks.decls.ok_for_sibcall (fndecl))
2894 maybe_complain_about_tail_call (exp, "frontend does not support"
2895 " sibling call");
2896 return false;
2899 /* All checks passed. */
2900 return true;
2903 /* Generate all the code for a CALL_EXPR exp
2904 and return an rtx for its value.
2905 Store the value in TARGET (specified as an rtx) if convenient.
2906 If the value is stored in TARGET then TARGET is returned.
2907 If IGNORE is nonzero, then we ignore the value of the function call. */
2910 expand_call (tree exp, rtx target, int ignore)
2912 /* Nonzero if we are currently expanding a call. */
2913 static int currently_expanding_call = 0;
2915 /* RTX for the function to be called. */
2916 rtx funexp;
2917 /* Sequence of insns to perform a normal "call". */
2918 rtx_insn *normal_call_insns = NULL;
2919 /* Sequence of insns to perform a tail "call". */
2920 rtx_insn *tail_call_insns = NULL;
2921 /* Data type of the function. */
2922 tree funtype;
2923 tree type_arg_types;
2924 tree rettype;
2925 /* Declaration of the function being called,
2926 or 0 if the function is computed (not known by name). */
2927 tree fndecl = 0;
2928 /* The type of the function being called. */
2929 tree fntype;
2930 bool try_tail_call = CALL_EXPR_TAILCALL (exp);
2931 bool must_tail_call = CALL_EXPR_MUST_TAIL_CALL (exp);
2932 int pass;
2934 /* Register in which non-BLKmode value will be returned,
2935 or 0 if no value or if value is BLKmode. */
2936 rtx valreg;
2937 /* Register(s) in which bounds are returned. */
2938 rtx valbnd = NULL;
2939 /* Address where we should return a BLKmode value;
2940 0 if value not BLKmode. */
2941 rtx structure_value_addr = 0;
2942 /* Nonzero if that address is being passed by treating it as
2943 an extra, implicit first parameter. Otherwise,
2944 it is passed by being copied directly into struct_value_rtx. */
2945 int structure_value_addr_parm = 0;
2946 /* Holds the value of implicit argument for the struct value. */
2947 tree structure_value_addr_value = NULL_TREE;
2948 /* Size of aggregate value wanted, or zero if none wanted
2949 or if we are using the non-reentrant PCC calling convention
2950 or expecting the value in registers. */
2951 HOST_WIDE_INT struct_value_size = 0;
2952 /* Nonzero if called function returns an aggregate in memory PCC style,
2953 by returning the address of where to find it. */
2954 int pcc_struct_value = 0;
2955 rtx struct_value = 0;
2957 /* Number of actual parameters in this call, including struct value addr. */
2958 int num_actuals;
2959 /* Number of named args. Args after this are anonymous ones
2960 and they must all go on the stack. */
2961 int n_named_args;
2962 /* Number of complex actual arguments that need to be split. */
2963 int num_complex_actuals = 0;
2965 /* Vector of information about each argument.
2966 Arguments are numbered in the order they will be pushed,
2967 not the order they are written. */
2968 struct arg_data *args;
2970 /* Total size in bytes of all the stack-parms scanned so far. */
2971 struct args_size args_size;
2972 struct args_size adjusted_args_size;
2973 /* Size of arguments before any adjustments (such as rounding). */
2974 int unadjusted_args_size;
2975 /* Data on reg parms scanned so far. */
2976 CUMULATIVE_ARGS args_so_far_v;
2977 cumulative_args_t args_so_far;
2978 /* Nonzero if a reg parm has been scanned. */
2979 int reg_parm_seen;
2980 /* Nonzero if this is an indirect function call. */
2982 /* Nonzero if we must avoid push-insns in the args for this call.
2983 If stack space is allocated for register parameters, but not by the
2984 caller, then it is preallocated in the fixed part of the stack frame.
2985 So the entire argument block must then be preallocated (i.e., we
2986 ignore PUSH_ROUNDING in that case). */
2988 int must_preallocate = !PUSH_ARGS;
2990 /* Size of the stack reserved for parameter registers. */
2991 int reg_parm_stack_space = 0;
2993 /* Address of space preallocated for stack parms
2994 (on machines that lack push insns), or 0 if space not preallocated. */
2995 rtx argblock = 0;
2997 /* Mask of ECF_ and ERF_ flags. */
2998 int flags = 0;
2999 int return_flags = 0;
3000 #ifdef REG_PARM_STACK_SPACE
3001 /* Define the boundary of the register parm stack space that needs to be
3002 saved, if any. */
3003 int low_to_save, high_to_save;
3004 rtx save_area = 0; /* Place that it is saved */
3005 #endif
3007 int initial_highest_arg_in_use = highest_outgoing_arg_in_use;
3008 char *initial_stack_usage_map = stack_usage_map;
3009 char *stack_usage_map_buf = NULL;
3011 int old_stack_allocated;
3013 /* State variables to track stack modifications. */
3014 rtx old_stack_level = 0;
3015 int old_stack_arg_under_construction = 0;
3016 int old_pending_adj = 0;
3017 int old_inhibit_defer_pop = inhibit_defer_pop;
3019 /* Some stack pointer alterations we make are performed via
3020 allocate_dynamic_stack_space. This modifies the stack_pointer_delta,
3021 which we then also need to save/restore along the way. */
3022 int old_stack_pointer_delta = 0;
3024 rtx call_fusage;
3025 tree addr = CALL_EXPR_FN (exp);
3026 int i;
3027 /* The alignment of the stack, in bits. */
3028 unsigned HOST_WIDE_INT preferred_stack_boundary;
3029 /* The alignment of the stack, in bytes. */
3030 unsigned HOST_WIDE_INT preferred_unit_stack_boundary;
3031 /* The static chain value to use for this call. */
3032 rtx static_chain_value;
3033 /* See if this is "nothrow" function call. */
3034 if (TREE_NOTHROW (exp))
3035 flags |= ECF_NOTHROW;
3037 /* See if we can find a DECL-node for the actual function, and get the
3038 function attributes (flags) from the function decl or type node. */
3039 fndecl = get_callee_fndecl (exp);
3040 if (fndecl)
3042 fntype = TREE_TYPE (fndecl);
3043 flags |= flags_from_decl_or_type (fndecl);
3044 return_flags |= decl_return_flags (fndecl);
3046 else
3048 fntype = TREE_TYPE (TREE_TYPE (addr));
3049 flags |= flags_from_decl_or_type (fntype);
3050 if (CALL_EXPR_BY_DESCRIPTOR (exp))
3051 flags |= ECF_BY_DESCRIPTOR;
3053 rettype = TREE_TYPE (exp);
3055 struct_value = targetm.calls.struct_value_rtx (fntype, 0);
3057 /* Warn if this value is an aggregate type,
3058 regardless of which calling convention we are using for it. */
3059 if (AGGREGATE_TYPE_P (rettype))
3060 warning (OPT_Waggregate_return, "function call has aggregate value");
3062 /* If the result of a non looping pure or const function call is
3063 ignored (or void), and none of its arguments are volatile, we can
3064 avoid expanding the call and just evaluate the arguments for
3065 side-effects. */
3066 if ((flags & (ECF_CONST | ECF_PURE))
3067 && (!(flags & ECF_LOOPING_CONST_OR_PURE))
3068 && (ignore || target == const0_rtx
3069 || TYPE_MODE (rettype) == VOIDmode))
3071 bool volatilep = false;
3072 tree arg;
3073 call_expr_arg_iterator iter;
3075 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
3076 if (TREE_THIS_VOLATILE (arg))
3078 volatilep = true;
3079 break;
3082 if (! volatilep)
3084 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
3085 expand_expr (arg, const0_rtx, VOIDmode, EXPAND_NORMAL);
3086 return const0_rtx;
3090 #ifdef REG_PARM_STACK_SPACE
3091 reg_parm_stack_space = REG_PARM_STACK_SPACE (!fndecl ? fntype : fndecl);
3092 #endif
3094 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))
3095 && reg_parm_stack_space > 0 && PUSH_ARGS)
3096 must_preallocate = 1;
3098 /* Set up a place to return a structure. */
3100 /* Cater to broken compilers. */
3101 if (aggregate_value_p (exp, fntype))
3103 /* This call returns a big structure. */
3104 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
3106 #ifdef PCC_STATIC_STRUCT_RETURN
3108 pcc_struct_value = 1;
3110 #else /* not PCC_STATIC_STRUCT_RETURN */
3112 struct_value_size = int_size_in_bytes (rettype);
3114 /* Even if it is semantically safe to use the target as the return
3115 slot, it may be not sufficiently aligned for the return type. */
3116 if (CALL_EXPR_RETURN_SLOT_OPT (exp)
3117 && target
3118 && MEM_P (target)
3119 && !(MEM_ALIGN (target) < TYPE_ALIGN (rettype)
3120 && targetm.slow_unaligned_access (TYPE_MODE (rettype),
3121 MEM_ALIGN (target))))
3122 structure_value_addr = XEXP (target, 0);
3123 else
3125 /* For variable-sized objects, we must be called with a target
3126 specified. If we were to allocate space on the stack here,
3127 we would have no way of knowing when to free it. */
3128 rtx d = assign_temp (rettype, 1, 1);
3129 structure_value_addr = XEXP (d, 0);
3130 target = 0;
3133 #endif /* not PCC_STATIC_STRUCT_RETURN */
3136 /* Figure out the amount to which the stack should be aligned. */
3137 preferred_stack_boundary = PREFERRED_STACK_BOUNDARY;
3138 if (fndecl)
3140 struct cgraph_rtl_info *i = cgraph_node::rtl_info (fndecl);
3141 /* Without automatic stack alignment, we can't increase preferred
3142 stack boundary. With automatic stack alignment, it is
3143 unnecessary since unless we can guarantee that all callers will
3144 align the outgoing stack properly, callee has to align its
3145 stack anyway. */
3146 if (i
3147 && i->preferred_incoming_stack_boundary
3148 && i->preferred_incoming_stack_boundary < preferred_stack_boundary)
3149 preferred_stack_boundary = i->preferred_incoming_stack_boundary;
3152 /* Operand 0 is a pointer-to-function; get the type of the function. */
3153 funtype = TREE_TYPE (addr);
3154 gcc_assert (POINTER_TYPE_P (funtype));
3155 funtype = TREE_TYPE (funtype);
3157 /* Count whether there are actual complex arguments that need to be split
3158 into their real and imaginary parts. Munge the type_arg_types
3159 appropriately here as well. */
3160 if (targetm.calls.split_complex_arg)
3162 call_expr_arg_iterator iter;
3163 tree arg;
3164 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
3166 tree type = TREE_TYPE (arg);
3167 if (type && TREE_CODE (type) == COMPLEX_TYPE
3168 && targetm.calls.split_complex_arg (type))
3169 num_complex_actuals++;
3171 type_arg_types = split_complex_types (TYPE_ARG_TYPES (funtype));
3173 else
3174 type_arg_types = TYPE_ARG_TYPES (funtype);
3176 if (flags & ECF_MAY_BE_ALLOCA)
3177 cfun->calls_alloca = 1;
3179 /* If struct_value_rtx is 0, it means pass the address
3180 as if it were an extra parameter. Put the argument expression
3181 in structure_value_addr_value. */
3182 if (structure_value_addr && struct_value == 0)
3184 /* If structure_value_addr is a REG other than
3185 virtual_outgoing_args_rtx, we can use always use it. If it
3186 is not a REG, we must always copy it into a register.
3187 If it is virtual_outgoing_args_rtx, we must copy it to another
3188 register in some cases. */
3189 rtx temp = (!REG_P (structure_value_addr)
3190 || (ACCUMULATE_OUTGOING_ARGS
3191 && stack_arg_under_construction
3192 && structure_value_addr == virtual_outgoing_args_rtx)
3193 ? copy_addr_to_reg (convert_memory_address
3194 (Pmode, structure_value_addr))
3195 : structure_value_addr);
3197 structure_value_addr_value =
3198 make_tree (build_pointer_type (TREE_TYPE (funtype)), temp);
3199 structure_value_addr_parm = CALL_WITH_BOUNDS_P (exp) ? 2 : 1;
3202 /* Count the arguments and set NUM_ACTUALS. */
3203 num_actuals =
3204 call_expr_nargs (exp) + num_complex_actuals + structure_value_addr_parm;
3206 /* Compute number of named args.
3207 First, do a raw count of the args for INIT_CUMULATIVE_ARGS. */
3209 if (type_arg_types != 0)
3210 n_named_args
3211 = (list_length (type_arg_types)
3212 /* Count the struct value address, if it is passed as a parm. */
3213 + structure_value_addr_parm);
3214 else
3215 /* If we know nothing, treat all args as named. */
3216 n_named_args = num_actuals;
3218 /* Start updating where the next arg would go.
3220 On some machines (such as the PA) indirect calls have a different
3221 calling convention than normal calls. The fourth argument in
3222 INIT_CUMULATIVE_ARGS tells the backend if this is an indirect call
3223 or not. */
3224 INIT_CUMULATIVE_ARGS (args_so_far_v, funtype, NULL_RTX, fndecl, n_named_args);
3225 args_so_far = pack_cumulative_args (&args_so_far_v);
3227 /* Now possibly adjust the number of named args.
3228 Normally, don't include the last named arg if anonymous args follow.
3229 We do include the last named arg if
3230 targetm.calls.strict_argument_naming() returns nonzero.
3231 (If no anonymous args follow, the result of list_length is actually
3232 one too large. This is harmless.)
3234 If targetm.calls.pretend_outgoing_varargs_named() returns
3235 nonzero, and targetm.calls.strict_argument_naming() returns zero,
3236 this machine will be able to place unnamed args that were passed
3237 in registers into the stack. So treat all args as named. This
3238 allows the insns emitting for a specific argument list to be
3239 independent of the function declaration.
3241 If targetm.calls.pretend_outgoing_varargs_named() returns zero,
3242 we do not have any reliable way to pass unnamed args in
3243 registers, so we must force them into memory. */
3245 if (type_arg_types != 0
3246 && targetm.calls.strict_argument_naming (args_so_far))
3248 else if (type_arg_types != 0
3249 && ! targetm.calls.pretend_outgoing_varargs_named (args_so_far))
3250 /* Don't include the last named arg. */
3251 --n_named_args;
3252 else
3253 /* Treat all args as named. */
3254 n_named_args = num_actuals;
3256 /* Make a vector to hold all the information about each arg. */
3257 args = XCNEWVEC (struct arg_data, num_actuals);
3259 /* Build up entries in the ARGS array, compute the size of the
3260 arguments into ARGS_SIZE, etc. */
3261 initialize_argument_information (num_actuals, args, &args_size,
3262 n_named_args, exp,
3263 structure_value_addr_value, fndecl, fntype,
3264 args_so_far, reg_parm_stack_space,
3265 &old_stack_level, &old_pending_adj,
3266 &must_preallocate, &flags,
3267 &try_tail_call, CALL_FROM_THUNK_P (exp));
3269 if (args_size.var)
3270 must_preallocate = 1;
3272 /* Now make final decision about preallocating stack space. */
3273 must_preallocate = finalize_must_preallocate (must_preallocate,
3274 num_actuals, args,
3275 &args_size);
3277 /* If the structure value address will reference the stack pointer, we
3278 must stabilize it. We don't need to do this if we know that we are
3279 not going to adjust the stack pointer in processing this call. */
3281 if (structure_value_addr
3282 && (reg_mentioned_p (virtual_stack_dynamic_rtx, structure_value_addr)
3283 || reg_mentioned_p (virtual_outgoing_args_rtx,
3284 structure_value_addr))
3285 && (args_size.var
3286 || (!ACCUMULATE_OUTGOING_ARGS && args_size.constant)))
3287 structure_value_addr = copy_to_reg (structure_value_addr);
3289 /* Tail calls can make things harder to debug, and we've traditionally
3290 pushed these optimizations into -O2. Don't try if we're already
3291 expanding a call, as that means we're an argument. Don't try if
3292 there's cleanups, as we know there's code to follow the call. */
3294 if (currently_expanding_call++ != 0
3295 || !flag_optimize_sibling_calls
3296 || args_size.var
3297 || dbg_cnt (tail_call) == false)
3298 try_tail_call = 0;
3300 /* If the user has marked the function as requiring tail-call
3301 optimization, attempt it. */
3302 if (must_tail_call)
3303 try_tail_call = 1;
3305 /* Rest of purposes for tail call optimizations to fail. */
3306 if (try_tail_call)
3307 try_tail_call = can_implement_as_sibling_call_p (exp,
3308 structure_value_addr,
3309 funtype,
3310 reg_parm_stack_space,
3311 fndecl,
3312 flags, addr, args_size);
3314 /* Check if caller and callee disagree in promotion of function
3315 return value. */
3316 if (try_tail_call)
3318 machine_mode caller_mode, caller_promoted_mode;
3319 machine_mode callee_mode, callee_promoted_mode;
3320 int caller_unsignedp, callee_unsignedp;
3321 tree caller_res = DECL_RESULT (current_function_decl);
3323 caller_unsignedp = TYPE_UNSIGNED (TREE_TYPE (caller_res));
3324 caller_mode = DECL_MODE (caller_res);
3325 callee_unsignedp = TYPE_UNSIGNED (TREE_TYPE (funtype));
3326 callee_mode = TYPE_MODE (TREE_TYPE (funtype));
3327 caller_promoted_mode
3328 = promote_function_mode (TREE_TYPE (caller_res), caller_mode,
3329 &caller_unsignedp,
3330 TREE_TYPE (current_function_decl), 1);
3331 callee_promoted_mode
3332 = promote_function_mode (TREE_TYPE (funtype), callee_mode,
3333 &callee_unsignedp,
3334 funtype, 1);
3335 if (caller_mode != VOIDmode
3336 && (caller_promoted_mode != callee_promoted_mode
3337 || ((caller_mode != caller_promoted_mode
3338 || callee_mode != callee_promoted_mode)
3339 && (caller_unsignedp != callee_unsignedp
3340 || partial_subreg_p (caller_mode, callee_mode)))))
3342 try_tail_call = 0;
3343 maybe_complain_about_tail_call (exp,
3344 "caller and callee disagree in"
3345 " promotion of function"
3346 " return value");
3350 /* Ensure current function's preferred stack boundary is at least
3351 what we need. Stack alignment may also increase preferred stack
3352 boundary. */
3353 if (crtl->preferred_stack_boundary < preferred_stack_boundary)
3354 crtl->preferred_stack_boundary = preferred_stack_boundary;
3355 else
3356 preferred_stack_boundary = crtl->preferred_stack_boundary;
3358 preferred_unit_stack_boundary = preferred_stack_boundary / BITS_PER_UNIT;
3360 /* We want to make two insn chains; one for a sibling call, the other
3361 for a normal call. We will select one of the two chains after
3362 initial RTL generation is complete. */
3363 for (pass = try_tail_call ? 0 : 1; pass < 2; pass++)
3365 int sibcall_failure = 0;
3366 /* We want to emit any pending stack adjustments before the tail
3367 recursion "call". That way we know any adjustment after the tail
3368 recursion call can be ignored if we indeed use the tail
3369 call expansion. */
3370 saved_pending_stack_adjust save;
3371 rtx_insn *insns, *before_call, *after_args;
3372 rtx next_arg_reg;
3374 if (pass == 0)
3376 /* State variables we need to save and restore between
3377 iterations. */
3378 save_pending_stack_adjust (&save);
3380 if (pass)
3381 flags &= ~ECF_SIBCALL;
3382 else
3383 flags |= ECF_SIBCALL;
3385 /* Other state variables that we must reinitialize each time
3386 through the loop (that are not initialized by the loop itself). */
3387 argblock = 0;
3388 call_fusage = 0;
3390 /* Start a new sequence for the normal call case.
3392 From this point on, if the sibling call fails, we want to set
3393 sibcall_failure instead of continuing the loop. */
3394 start_sequence ();
3396 /* Don't let pending stack adjusts add up to too much.
3397 Also, do all pending adjustments now if there is any chance
3398 this might be a call to alloca or if we are expanding a sibling
3399 call sequence.
3400 Also do the adjustments before a throwing call, otherwise
3401 exception handling can fail; PR 19225. */
3402 if (pending_stack_adjust >= 32
3403 || (pending_stack_adjust > 0
3404 && (flags & ECF_MAY_BE_ALLOCA))
3405 || (pending_stack_adjust > 0
3406 && flag_exceptions && !(flags & ECF_NOTHROW))
3407 || pass == 0)
3408 do_pending_stack_adjust ();
3410 /* Precompute any arguments as needed. */
3411 if (pass)
3412 precompute_arguments (num_actuals, args);
3414 /* Now we are about to start emitting insns that can be deleted
3415 if a libcall is deleted. */
3416 if (pass && (flags & ECF_MALLOC))
3417 start_sequence ();
3419 if (pass == 0
3420 && crtl->stack_protect_guard
3421 && targetm.stack_protect_runtime_enabled_p ())
3422 stack_protect_epilogue ();
3424 adjusted_args_size = args_size;
3425 /* Compute the actual size of the argument block required. The variable
3426 and constant sizes must be combined, the size may have to be rounded,
3427 and there may be a minimum required size. When generating a sibcall
3428 pattern, do not round up, since we'll be re-using whatever space our
3429 caller provided. */
3430 unadjusted_args_size
3431 = compute_argument_block_size (reg_parm_stack_space,
3432 &adjusted_args_size,
3433 fndecl, fntype,
3434 (pass == 0 ? 0
3435 : preferred_stack_boundary));
3437 old_stack_allocated = stack_pointer_delta - pending_stack_adjust;
3439 /* The argument block when performing a sibling call is the
3440 incoming argument block. */
3441 if (pass == 0)
3443 argblock = crtl->args.internal_arg_pointer;
3444 if (STACK_GROWS_DOWNWARD)
3445 argblock
3446 = plus_constant (Pmode, argblock, crtl->args.pretend_args_size);
3447 else
3448 argblock
3449 = plus_constant (Pmode, argblock, -crtl->args.pretend_args_size);
3451 stored_args_map = sbitmap_alloc (args_size.constant);
3452 bitmap_clear (stored_args_map);
3455 /* If we have no actual push instructions, or shouldn't use them,
3456 make space for all args right now. */
3457 else if (adjusted_args_size.var != 0)
3459 if (old_stack_level == 0)
3461 emit_stack_save (SAVE_BLOCK, &old_stack_level);
3462 old_stack_pointer_delta = stack_pointer_delta;
3463 old_pending_adj = pending_stack_adjust;
3464 pending_stack_adjust = 0;
3465 /* stack_arg_under_construction says whether a stack arg is
3466 being constructed at the old stack level. Pushing the stack
3467 gets a clean outgoing argument block. */
3468 old_stack_arg_under_construction = stack_arg_under_construction;
3469 stack_arg_under_construction = 0;
3471 argblock = push_block (ARGS_SIZE_RTX (adjusted_args_size), 0, 0);
3472 if (flag_stack_usage_info)
3473 current_function_has_unbounded_dynamic_stack_size = 1;
3475 else
3477 /* Note that we must go through the motions of allocating an argument
3478 block even if the size is zero because we may be storing args
3479 in the area reserved for register arguments, which may be part of
3480 the stack frame. */
3482 int needed = adjusted_args_size.constant;
3484 /* Store the maximum argument space used. It will be pushed by
3485 the prologue (if ACCUMULATE_OUTGOING_ARGS, or stack overflow
3486 checking). */
3488 if (needed > crtl->outgoing_args_size)
3489 crtl->outgoing_args_size = needed;
3491 if (must_preallocate)
3493 if (ACCUMULATE_OUTGOING_ARGS)
3495 /* Since the stack pointer will never be pushed, it is
3496 possible for the evaluation of a parm to clobber
3497 something we have already written to the stack.
3498 Since most function calls on RISC machines do not use
3499 the stack, this is uncommon, but must work correctly.
3501 Therefore, we save any area of the stack that was already
3502 written and that we are using. Here we set up to do this
3503 by making a new stack usage map from the old one. The
3504 actual save will be done by store_one_arg.
3506 Another approach might be to try to reorder the argument
3507 evaluations to avoid this conflicting stack usage. */
3509 /* Since we will be writing into the entire argument area,
3510 the map must be allocated for its entire size, not just
3511 the part that is the responsibility of the caller. */
3512 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
3513 needed += reg_parm_stack_space;
3515 if (ARGS_GROW_DOWNWARD)
3516 highest_outgoing_arg_in_use
3517 = MAX (initial_highest_arg_in_use, needed + 1);
3518 else
3519 highest_outgoing_arg_in_use
3520 = MAX (initial_highest_arg_in_use, needed);
3522 free (stack_usage_map_buf);
3523 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use);
3524 stack_usage_map = stack_usage_map_buf;
3526 if (initial_highest_arg_in_use)
3527 memcpy (stack_usage_map, initial_stack_usage_map,
3528 initial_highest_arg_in_use);
3530 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use)
3531 memset (&stack_usage_map[initial_highest_arg_in_use], 0,
3532 (highest_outgoing_arg_in_use
3533 - initial_highest_arg_in_use));
3534 needed = 0;
3536 /* The address of the outgoing argument list must not be
3537 copied to a register here, because argblock would be left
3538 pointing to the wrong place after the call to
3539 allocate_dynamic_stack_space below. */
3541 argblock = virtual_outgoing_args_rtx;
3543 else
3545 if (inhibit_defer_pop == 0)
3547 /* Try to reuse some or all of the pending_stack_adjust
3548 to get this space. */
3549 needed
3550 = (combine_pending_stack_adjustment_and_call
3551 (unadjusted_args_size,
3552 &adjusted_args_size,
3553 preferred_unit_stack_boundary));
3555 /* combine_pending_stack_adjustment_and_call computes
3556 an adjustment before the arguments are allocated.
3557 Account for them and see whether or not the stack
3558 needs to go up or down. */
3559 needed = unadjusted_args_size - needed;
3561 if (needed < 0)
3563 /* We're releasing stack space. */
3564 /* ??? We can avoid any adjustment at all if we're
3565 already aligned. FIXME. */
3566 pending_stack_adjust = -needed;
3567 do_pending_stack_adjust ();
3568 needed = 0;
3570 else
3571 /* We need to allocate space. We'll do that in
3572 push_block below. */
3573 pending_stack_adjust = 0;
3576 /* Special case this because overhead of `push_block' in
3577 this case is non-trivial. */
3578 if (needed == 0)
3579 argblock = virtual_outgoing_args_rtx;
3580 else
3582 argblock = push_block (GEN_INT (needed), 0, 0);
3583 if (ARGS_GROW_DOWNWARD)
3584 argblock = plus_constant (Pmode, argblock, needed);
3587 /* We only really need to call `copy_to_reg' in the case
3588 where push insns are going to be used to pass ARGBLOCK
3589 to a function call in ARGS. In that case, the stack
3590 pointer changes value from the allocation point to the
3591 call point, and hence the value of
3592 VIRTUAL_OUTGOING_ARGS_RTX changes as well. But might
3593 as well always do it. */
3594 argblock = copy_to_reg (argblock);
3599 if (ACCUMULATE_OUTGOING_ARGS)
3601 /* The save/restore code in store_one_arg handles all
3602 cases except one: a constructor call (including a C
3603 function returning a BLKmode struct) to initialize
3604 an argument. */
3605 if (stack_arg_under_construction)
3607 rtx push_size
3608 = GEN_INT (adjusted_args_size.constant
3609 + (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype
3610 : TREE_TYPE (fndecl))) ? 0
3611 : reg_parm_stack_space));
3612 if (old_stack_level == 0)
3614 emit_stack_save (SAVE_BLOCK, &old_stack_level);
3615 old_stack_pointer_delta = stack_pointer_delta;
3616 old_pending_adj = pending_stack_adjust;
3617 pending_stack_adjust = 0;
3618 /* stack_arg_under_construction says whether a stack
3619 arg is being constructed at the old stack level.
3620 Pushing the stack gets a clean outgoing argument
3621 block. */
3622 old_stack_arg_under_construction
3623 = stack_arg_under_construction;
3624 stack_arg_under_construction = 0;
3625 /* Make a new map for the new argument list. */
3626 free (stack_usage_map_buf);
3627 stack_usage_map_buf = XCNEWVEC (char, highest_outgoing_arg_in_use);
3628 stack_usage_map = stack_usage_map_buf;
3629 highest_outgoing_arg_in_use = 0;
3631 /* We can pass TRUE as the 4th argument because we just
3632 saved the stack pointer and will restore it right after
3633 the call. */
3634 allocate_dynamic_stack_space (push_size, 0, BIGGEST_ALIGNMENT,
3635 -1, true);
3638 /* If argument evaluation might modify the stack pointer,
3639 copy the address of the argument list to a register. */
3640 for (i = 0; i < num_actuals; i++)
3641 if (args[i].pass_on_stack)
3643 argblock = copy_addr_to_reg (argblock);
3644 break;
3648 compute_argument_addresses (args, argblock, num_actuals);
3650 /* Stack is properly aligned, pops can't safely be deferred during
3651 the evaluation of the arguments. */
3652 NO_DEFER_POP;
3654 /* Precompute all register parameters. It isn't safe to compute
3655 anything once we have started filling any specific hard regs.
3656 TLS symbols sometimes need a call to resolve. Precompute
3657 register parameters before any stack pointer manipulation
3658 to avoid unaligned stack in the called function. */
3659 precompute_register_parameters (num_actuals, args, &reg_parm_seen);
3661 OK_DEFER_POP;
3663 /* Perform stack alignment before the first push (the last arg). */
3664 if (argblock == 0
3665 && adjusted_args_size.constant > reg_parm_stack_space
3666 && adjusted_args_size.constant != unadjusted_args_size)
3668 /* When the stack adjustment is pending, we get better code
3669 by combining the adjustments. */
3670 if (pending_stack_adjust
3671 && ! inhibit_defer_pop)
3673 pending_stack_adjust
3674 = (combine_pending_stack_adjustment_and_call
3675 (unadjusted_args_size,
3676 &adjusted_args_size,
3677 preferred_unit_stack_boundary));
3678 do_pending_stack_adjust ();
3680 else if (argblock == 0)
3681 anti_adjust_stack (GEN_INT (adjusted_args_size.constant
3682 - unadjusted_args_size));
3684 /* Now that the stack is properly aligned, pops can't safely
3685 be deferred during the evaluation of the arguments. */
3686 NO_DEFER_POP;
3688 /* Record the maximum pushed stack space size. We need to delay
3689 doing it this far to take into account the optimization done
3690 by combine_pending_stack_adjustment_and_call. */
3691 if (flag_stack_usage_info
3692 && !ACCUMULATE_OUTGOING_ARGS
3693 && pass
3694 && adjusted_args_size.var == 0)
3696 int pushed = adjusted_args_size.constant + pending_stack_adjust;
3697 if (pushed > current_function_pushed_stack_size)
3698 current_function_pushed_stack_size = pushed;
3701 funexp = rtx_for_function_call (fndecl, addr);
3703 if (CALL_EXPR_STATIC_CHAIN (exp))
3704 static_chain_value = expand_normal (CALL_EXPR_STATIC_CHAIN (exp));
3705 else
3706 static_chain_value = 0;
3708 #ifdef REG_PARM_STACK_SPACE
3709 /* Save the fixed argument area if it's part of the caller's frame and
3710 is clobbered by argument setup for this call. */
3711 if (ACCUMULATE_OUTGOING_ARGS && pass)
3712 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock,
3713 &low_to_save, &high_to_save);
3714 #endif
3716 /* Now store (and compute if necessary) all non-register parms.
3717 These come before register parms, since they can require block-moves,
3718 which could clobber the registers used for register parms.
3719 Parms which have partial registers are not stored here,
3720 but we do preallocate space here if they want that. */
3722 for (i = 0; i < num_actuals; i++)
3724 /* Delay bounds until all other args are stored. */
3725 if (POINTER_BOUNDS_P (args[i].tree_value))
3726 continue;
3727 else if (args[i].reg == 0 || args[i].pass_on_stack)
3729 rtx_insn *before_arg = get_last_insn ();
3731 /* We don't allow passing huge (> 2^30 B) arguments
3732 by value. It would cause an overflow later on. */
3733 if (adjusted_args_size.constant
3734 >= (1 << (HOST_BITS_PER_INT - 2)))
3736 sorry ("passing too large argument on stack");
3737 continue;
3740 if (store_one_arg (&args[i], argblock, flags,
3741 adjusted_args_size.var != 0,
3742 reg_parm_stack_space)
3743 || (pass == 0
3744 && check_sibcall_argument_overlap (before_arg,
3745 &args[i], 1)))
3746 sibcall_failure = 1;
3749 if (args[i].stack)
3750 call_fusage
3751 = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[i].tree_value)),
3752 gen_rtx_USE (VOIDmode, args[i].stack),
3753 call_fusage);
3756 /* If we have a parm that is passed in registers but not in memory
3757 and whose alignment does not permit a direct copy into registers,
3758 make a group of pseudos that correspond to each register that we
3759 will later fill. */
3760 if (STRICT_ALIGNMENT)
3761 store_unaligned_arguments_into_pseudos (args, num_actuals);
3763 /* Now store any partially-in-registers parm.
3764 This is the last place a block-move can happen. */
3765 if (reg_parm_seen)
3766 for (i = 0; i < num_actuals; i++)
3767 if (args[i].partial != 0 && ! args[i].pass_on_stack)
3769 rtx_insn *before_arg = get_last_insn ();
3771 /* On targets with weird calling conventions (e.g. PA) it's
3772 hard to ensure that all cases of argument overlap between
3773 stack and registers work. Play it safe and bail out. */
3774 if (ARGS_GROW_DOWNWARD && !STACK_GROWS_DOWNWARD)
3776 sibcall_failure = 1;
3777 break;
3780 if (store_one_arg (&args[i], argblock, flags,
3781 adjusted_args_size.var != 0,
3782 reg_parm_stack_space)
3783 || (pass == 0
3784 && check_sibcall_argument_overlap (before_arg,
3785 &args[i], 1)))
3786 sibcall_failure = 1;
3789 bool any_regs = false;
3790 for (i = 0; i < num_actuals; i++)
3791 if (args[i].reg != NULL_RTX)
3793 any_regs = true;
3794 targetm.calls.call_args (args[i].reg, funtype);
3796 if (!any_regs)
3797 targetm.calls.call_args (pc_rtx, funtype);
3799 /* Figure out the register where the value, if any, will come back. */
3800 valreg = 0;
3801 valbnd = 0;
3802 if (TYPE_MODE (rettype) != VOIDmode
3803 && ! structure_value_addr)
3805 if (pcc_struct_value)
3807 valreg = hard_function_value (build_pointer_type (rettype),
3808 fndecl, NULL, (pass == 0));
3809 if (CALL_WITH_BOUNDS_P (exp))
3810 valbnd = targetm.calls.
3811 chkp_function_value_bounds (build_pointer_type (rettype),
3812 fndecl, (pass == 0));
3814 else
3816 valreg = hard_function_value (rettype, fndecl, fntype,
3817 (pass == 0));
3818 if (CALL_WITH_BOUNDS_P (exp))
3819 valbnd = targetm.calls.chkp_function_value_bounds (rettype,
3820 fndecl,
3821 (pass == 0));
3824 /* If VALREG is a PARALLEL whose first member has a zero
3825 offset, use that. This is for targets such as m68k that
3826 return the same value in multiple places. */
3827 if (GET_CODE (valreg) == PARALLEL)
3829 rtx elem = XVECEXP (valreg, 0, 0);
3830 rtx where = XEXP (elem, 0);
3831 rtx offset = XEXP (elem, 1);
3832 if (offset == const0_rtx
3833 && GET_MODE (where) == GET_MODE (valreg))
3834 valreg = where;
3838 /* Store all bounds not passed in registers. */
3839 for (i = 0; i < num_actuals; i++)
3841 if (POINTER_BOUNDS_P (args[i].tree_value)
3842 && !args[i].reg)
3843 store_bounds (&args[i],
3844 args[i].pointer_arg == -1
3845 ? NULL
3846 : &args[args[i].pointer_arg]);
3849 /* If register arguments require space on the stack and stack space
3850 was not preallocated, allocate stack space here for arguments
3851 passed in registers. */
3852 if (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))
3853 && !ACCUMULATE_OUTGOING_ARGS
3854 && must_preallocate == 0 && reg_parm_stack_space > 0)
3855 anti_adjust_stack (GEN_INT (reg_parm_stack_space));
3857 /* Pass the function the address in which to return a
3858 structure value. */
3859 if (pass != 0 && structure_value_addr && ! structure_value_addr_parm)
3861 structure_value_addr
3862 = convert_memory_address (Pmode, structure_value_addr);
3863 emit_move_insn (struct_value,
3864 force_reg (Pmode,
3865 force_operand (structure_value_addr,
3866 NULL_RTX)));
3868 if (REG_P (struct_value))
3869 use_reg (&call_fusage, struct_value);
3872 after_args = get_last_insn ();
3873 funexp = prepare_call_address (fndecl ? fndecl : fntype, funexp,
3874 static_chain_value, &call_fusage,
3875 reg_parm_seen, flags);
3877 load_register_parameters (args, num_actuals, &call_fusage, flags,
3878 pass == 0, &sibcall_failure);
3880 /* Save a pointer to the last insn before the call, so that we can
3881 later safely search backwards to find the CALL_INSN. */
3882 before_call = get_last_insn ();
3884 /* Set up next argument register. For sibling calls on machines
3885 with register windows this should be the incoming register. */
3886 if (pass == 0)
3887 next_arg_reg = targetm.calls.function_incoming_arg (args_so_far,
3888 VOIDmode,
3889 void_type_node,
3890 true);
3891 else
3892 next_arg_reg = targetm.calls.function_arg (args_so_far,
3893 VOIDmode, void_type_node,
3894 true);
3896 if (pass == 1 && (return_flags & ERF_RETURNS_ARG))
3898 int arg_nr = return_flags & ERF_RETURN_ARG_MASK;
3899 arg_nr = num_actuals - arg_nr - 1;
3900 if (arg_nr >= 0
3901 && arg_nr < num_actuals
3902 && args[arg_nr].reg
3903 && valreg
3904 && REG_P (valreg)
3905 && GET_MODE (args[arg_nr].reg) == GET_MODE (valreg))
3906 call_fusage
3907 = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[arg_nr].tree_value)),
3908 gen_rtx_SET (valreg, args[arg_nr].reg),
3909 call_fusage);
3911 /* All arguments and registers used for the call must be set up by
3912 now! */
3914 /* Stack must be properly aligned now. */
3915 gcc_assert (!pass
3916 || !(stack_pointer_delta % preferred_unit_stack_boundary));
3918 /* Generate the actual call instruction. */
3919 emit_call_1 (funexp, exp, fndecl, funtype, unadjusted_args_size,
3920 adjusted_args_size.constant, struct_value_size,
3921 next_arg_reg, valreg, old_inhibit_defer_pop, call_fusage,
3922 flags, args_so_far);
3924 if (flag_ipa_ra)
3926 rtx_call_insn *last;
3927 rtx datum = NULL_RTX;
3928 if (fndecl != NULL_TREE)
3930 datum = XEXP (DECL_RTL (fndecl), 0);
3931 gcc_assert (datum != NULL_RTX
3932 && GET_CODE (datum) == SYMBOL_REF);
3934 last = last_call_insn ();
3935 add_reg_note (last, REG_CALL_DECL, datum);
3938 /* If the call setup or the call itself overlaps with anything
3939 of the argument setup we probably clobbered our call address.
3940 In that case we can't do sibcalls. */
3941 if (pass == 0
3942 && check_sibcall_argument_overlap (after_args, 0, 0))
3943 sibcall_failure = 1;
3945 /* If a non-BLKmode value is returned at the most significant end
3946 of a register, shift the register right by the appropriate amount
3947 and update VALREG accordingly. BLKmode values are handled by the
3948 group load/store machinery below. */
3949 if (!structure_value_addr
3950 && !pcc_struct_value
3951 && TYPE_MODE (rettype) != VOIDmode
3952 && TYPE_MODE (rettype) != BLKmode
3953 && REG_P (valreg)
3954 && targetm.calls.return_in_msb (rettype))
3956 if (shift_return_value (TYPE_MODE (rettype), false, valreg))
3957 sibcall_failure = 1;
3958 valreg = gen_rtx_REG (TYPE_MODE (rettype), REGNO (valreg));
3961 if (pass && (flags & ECF_MALLOC))
3963 rtx temp = gen_reg_rtx (GET_MODE (valreg));
3964 rtx_insn *last, *insns;
3966 /* The return value from a malloc-like function is a pointer. */
3967 if (TREE_CODE (rettype) == POINTER_TYPE)
3968 mark_reg_pointer (temp, MALLOC_ABI_ALIGNMENT);
3970 emit_move_insn (temp, valreg);
3972 /* The return value from a malloc-like function can not alias
3973 anything else. */
3974 last = get_last_insn ();
3975 add_reg_note (last, REG_NOALIAS, temp);
3977 /* Write out the sequence. */
3978 insns = get_insns ();
3979 end_sequence ();
3980 emit_insn (insns);
3981 valreg = temp;
3984 /* For calls to `setjmp', etc., inform
3985 function.c:setjmp_warnings that it should complain if
3986 nonvolatile values are live. For functions that cannot
3987 return, inform flow that control does not fall through. */
3989 if ((flags & ECF_NORETURN) || pass == 0)
3991 /* The barrier must be emitted
3992 immediately after the CALL_INSN. Some ports emit more
3993 than just a CALL_INSN above, so we must search for it here. */
3995 rtx_insn *last = get_last_insn ();
3996 while (!CALL_P (last))
3998 last = PREV_INSN (last);
3999 /* There was no CALL_INSN? */
4000 gcc_assert (last != before_call);
4003 emit_barrier_after (last);
4005 /* Stack adjustments after a noreturn call are dead code.
4006 However when NO_DEFER_POP is in effect, we must preserve
4007 stack_pointer_delta. */
4008 if (inhibit_defer_pop == 0)
4010 stack_pointer_delta = old_stack_allocated;
4011 pending_stack_adjust = 0;
4015 /* If value type not void, return an rtx for the value. */
4017 if (TYPE_MODE (rettype) == VOIDmode
4018 || ignore)
4019 target = const0_rtx;
4020 else if (structure_value_addr)
4022 if (target == 0 || !MEM_P (target))
4024 target
4025 = gen_rtx_MEM (TYPE_MODE (rettype),
4026 memory_address (TYPE_MODE (rettype),
4027 structure_value_addr));
4028 set_mem_attributes (target, rettype, 1);
4031 else if (pcc_struct_value)
4033 /* This is the special C++ case where we need to
4034 know what the true target was. We take care to
4035 never use this value more than once in one expression. */
4036 target = gen_rtx_MEM (TYPE_MODE (rettype),
4037 copy_to_reg (valreg));
4038 set_mem_attributes (target, rettype, 1);
4040 /* Handle calls that return values in multiple non-contiguous locations.
4041 The Irix 6 ABI has examples of this. */
4042 else if (GET_CODE (valreg) == PARALLEL)
4044 if (target == 0)
4045 target = emit_group_move_into_temps (valreg);
4046 else if (rtx_equal_p (target, valreg))
4048 else if (GET_CODE (target) == PARALLEL)
4049 /* Handle the result of a emit_group_move_into_temps
4050 call in the previous pass. */
4051 emit_group_move (target, valreg);
4052 else
4053 emit_group_store (target, valreg, rettype,
4054 int_size_in_bytes (rettype));
4056 else if (target
4057 && GET_MODE (target) == TYPE_MODE (rettype)
4058 && GET_MODE (target) == GET_MODE (valreg))
4060 bool may_overlap = false;
4062 /* We have to copy a return value in a CLASS_LIKELY_SPILLED hard
4063 reg to a plain register. */
4064 if (!REG_P (target) || HARD_REGISTER_P (target))
4065 valreg = avoid_likely_spilled_reg (valreg);
4067 /* If TARGET is a MEM in the argument area, and we have
4068 saved part of the argument area, then we can't store
4069 directly into TARGET as it may get overwritten when we
4070 restore the argument save area below. Don't work too
4071 hard though and simply force TARGET to a register if it
4072 is a MEM; the optimizer is quite likely to sort it out. */
4073 if (ACCUMULATE_OUTGOING_ARGS && pass && MEM_P (target))
4074 for (i = 0; i < num_actuals; i++)
4075 if (args[i].save_area)
4077 may_overlap = true;
4078 break;
4081 if (may_overlap)
4082 target = copy_to_reg (valreg);
4083 else
4085 /* TARGET and VALREG cannot be equal at this point
4086 because the latter would not have
4087 REG_FUNCTION_VALUE_P true, while the former would if
4088 it were referring to the same register.
4090 If they refer to the same register, this move will be
4091 a no-op, except when function inlining is being
4092 done. */
4093 emit_move_insn (target, valreg);
4095 /* If we are setting a MEM, this code must be executed.
4096 Since it is emitted after the call insn, sibcall
4097 optimization cannot be performed in that case. */
4098 if (MEM_P (target))
4099 sibcall_failure = 1;
4102 else
4103 target = copy_to_reg (avoid_likely_spilled_reg (valreg));
4105 /* If we promoted this return value, make the proper SUBREG.
4106 TARGET might be const0_rtx here, so be careful. */
4107 if (REG_P (target)
4108 && TYPE_MODE (rettype) != BLKmode
4109 && GET_MODE (target) != TYPE_MODE (rettype))
4111 tree type = rettype;
4112 int unsignedp = TYPE_UNSIGNED (type);
4113 machine_mode pmode;
4115 /* Ensure we promote as expected, and get the new unsignedness. */
4116 pmode = promote_function_mode (type, TYPE_MODE (type), &unsignedp,
4117 funtype, 1);
4118 gcc_assert (GET_MODE (target) == pmode);
4120 unsigned int offset = subreg_lowpart_offset (TYPE_MODE (type),
4121 GET_MODE (target));
4122 target = gen_rtx_SUBREG (TYPE_MODE (type), target, offset);
4123 SUBREG_PROMOTED_VAR_P (target) = 1;
4124 SUBREG_PROMOTED_SET (target, unsignedp);
4127 /* If size of args is variable or this was a constructor call for a stack
4128 argument, restore saved stack-pointer value. */
4130 if (old_stack_level)
4132 rtx_insn *prev = get_last_insn ();
4134 emit_stack_restore (SAVE_BLOCK, old_stack_level);
4135 stack_pointer_delta = old_stack_pointer_delta;
4137 fixup_args_size_notes (prev, get_last_insn (), stack_pointer_delta);
4139 pending_stack_adjust = old_pending_adj;
4140 old_stack_allocated = stack_pointer_delta - pending_stack_adjust;
4141 stack_arg_under_construction = old_stack_arg_under_construction;
4142 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
4143 stack_usage_map = initial_stack_usage_map;
4144 sibcall_failure = 1;
4146 else if (ACCUMULATE_OUTGOING_ARGS && pass)
4148 #ifdef REG_PARM_STACK_SPACE
4149 if (save_area)
4150 restore_fixed_argument_area (save_area, argblock,
4151 high_to_save, low_to_save);
4152 #endif
4154 /* If we saved any argument areas, restore them. */
4155 for (i = 0; i < num_actuals; i++)
4156 if (args[i].save_area)
4158 machine_mode save_mode = GET_MODE (args[i].save_area);
4159 rtx stack_area
4160 = gen_rtx_MEM (save_mode,
4161 memory_address (save_mode,
4162 XEXP (args[i].stack_slot, 0)));
4164 if (save_mode != BLKmode)
4165 emit_move_insn (stack_area, args[i].save_area);
4166 else
4167 emit_block_move (stack_area, args[i].save_area,
4168 GEN_INT (args[i].locate.size.constant),
4169 BLOCK_OP_CALL_PARM);
4172 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
4173 stack_usage_map = initial_stack_usage_map;
4176 /* If this was alloca, record the new stack level. */
4177 if (flags & ECF_MAY_BE_ALLOCA)
4178 record_new_stack_level ();
4180 /* Free up storage we no longer need. */
4181 for (i = 0; i < num_actuals; ++i)
4182 free (args[i].aligned_regs);
4184 targetm.calls.end_call_args ();
4186 insns = get_insns ();
4187 end_sequence ();
4189 if (pass == 0)
4191 tail_call_insns = insns;
4193 /* Restore the pending stack adjustment now that we have
4194 finished generating the sibling call sequence. */
4196 restore_pending_stack_adjust (&save);
4198 /* Prepare arg structure for next iteration. */
4199 for (i = 0; i < num_actuals; i++)
4201 args[i].value = 0;
4202 args[i].aligned_regs = 0;
4203 args[i].stack = 0;
4206 sbitmap_free (stored_args_map);
4207 internal_arg_pointer_exp_state.scan_start = NULL;
4208 internal_arg_pointer_exp_state.cache.release ();
4210 else
4212 normal_call_insns = insns;
4214 /* Verify that we've deallocated all the stack we used. */
4215 gcc_assert ((flags & ECF_NORETURN)
4216 || (old_stack_allocated
4217 == stack_pointer_delta - pending_stack_adjust));
4220 /* If something prevents making this a sibling call,
4221 zero out the sequence. */
4222 if (sibcall_failure)
4223 tail_call_insns = NULL;
4224 else
4225 break;
4228 /* If tail call production succeeded, we need to remove REG_EQUIV notes on
4229 arguments too, as argument area is now clobbered by the call. */
4230 if (tail_call_insns)
4232 emit_insn (tail_call_insns);
4233 crtl->tail_call_emit = true;
4235 else
4237 emit_insn (normal_call_insns);
4238 if (try_tail_call)
4239 /* Ideally we'd emit a message for all of the ways that it could
4240 have failed. */
4241 maybe_complain_about_tail_call (exp, "tail call production failed");
4244 currently_expanding_call--;
4246 free (stack_usage_map_buf);
4247 free (args);
4249 /* Join result with returned bounds so caller may use them if needed. */
4250 target = chkp_join_splitted_slot (target, valbnd);
4252 return target;
4255 /* A sibling call sequence invalidates any REG_EQUIV notes made for
4256 this function's incoming arguments.
4258 At the start of RTL generation we know the only REG_EQUIV notes
4259 in the rtl chain are those for incoming arguments, so we can look
4260 for REG_EQUIV notes between the start of the function and the
4261 NOTE_INSN_FUNCTION_BEG.
4263 This is (slight) overkill. We could keep track of the highest
4264 argument we clobber and be more selective in removing notes, but it
4265 does not seem to be worth the effort. */
4267 void
4268 fixup_tail_calls (void)
4270 rtx_insn *insn;
4272 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
4274 rtx note;
4276 /* There are never REG_EQUIV notes for the incoming arguments
4277 after the NOTE_INSN_FUNCTION_BEG note, so stop if we see it. */
4278 if (NOTE_P (insn)
4279 && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG)
4280 break;
4282 note = find_reg_note (insn, REG_EQUIV, 0);
4283 if (note)
4284 remove_note (insn, note);
4285 note = find_reg_note (insn, REG_EQUIV, 0);
4286 gcc_assert (!note);
4290 /* Traverse a list of TYPES and expand all complex types into their
4291 components. */
4292 static tree
4293 split_complex_types (tree types)
4295 tree p;
4297 /* Before allocating memory, check for the common case of no complex. */
4298 for (p = types; p; p = TREE_CHAIN (p))
4300 tree type = TREE_VALUE (p);
4301 if (TREE_CODE (type) == COMPLEX_TYPE
4302 && targetm.calls.split_complex_arg (type))
4303 goto found;
4305 return types;
4307 found:
4308 types = copy_list (types);
4310 for (p = types; p; p = TREE_CHAIN (p))
4312 tree complex_type = TREE_VALUE (p);
4314 if (TREE_CODE (complex_type) == COMPLEX_TYPE
4315 && targetm.calls.split_complex_arg (complex_type))
4317 tree next, imag;
4319 /* Rewrite complex type with component type. */
4320 TREE_VALUE (p) = TREE_TYPE (complex_type);
4321 next = TREE_CHAIN (p);
4323 /* Add another component type for the imaginary part. */
4324 imag = build_tree_list (NULL_TREE, TREE_VALUE (p));
4325 TREE_CHAIN (p) = imag;
4326 TREE_CHAIN (imag) = next;
4328 /* Skip the newly created node. */
4329 p = TREE_CHAIN (p);
4333 return types;
4336 /* Output a library call to function ORGFUN (a SYMBOL_REF rtx)
4337 for a value of mode OUTMODE,
4338 with NARGS different arguments, passed as ARGS.
4339 Store the return value if RETVAL is nonzero: store it in VALUE if
4340 VALUE is nonnull, otherwise pick a convenient location. In either
4341 case return the location of the stored value.
4343 FN_TYPE should be LCT_NORMAL for `normal' calls, LCT_CONST for
4344 `const' calls, LCT_PURE for `pure' calls, or another LCT_ value for
4345 other types of library calls. */
4348 emit_library_call_value_1 (int retval, rtx orgfun, rtx value,
4349 enum libcall_type fn_type,
4350 machine_mode outmode, int nargs, rtx_mode_t *args)
4352 /* Total size in bytes of all the stack-parms scanned so far. */
4353 struct args_size args_size;
4354 /* Size of arguments before any adjustments (such as rounding). */
4355 struct args_size original_args_size;
4356 int argnum;
4357 rtx fun;
4358 /* Todo, choose the correct decl type of orgfun. Sadly this information
4359 isn't present here, so we default to native calling abi here. */
4360 tree fndecl ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */
4361 tree fntype ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */
4362 int count;
4363 rtx argblock = 0;
4364 CUMULATIVE_ARGS args_so_far_v;
4365 cumulative_args_t args_so_far;
4366 struct arg
4368 rtx value;
4369 machine_mode mode;
4370 rtx reg;
4371 int partial;
4372 struct locate_and_pad_arg_data locate;
4373 rtx save_area;
4375 struct arg *argvec;
4376 int old_inhibit_defer_pop = inhibit_defer_pop;
4377 rtx call_fusage = 0;
4378 rtx mem_value = 0;
4379 rtx valreg;
4380 int pcc_struct_value = 0;
4381 int struct_value_size = 0;
4382 int flags;
4383 int reg_parm_stack_space = 0;
4384 int needed;
4385 rtx_insn *before_call;
4386 bool have_push_fusage;
4387 tree tfom; /* type_for_mode (outmode, 0) */
4389 #ifdef REG_PARM_STACK_SPACE
4390 /* Define the boundary of the register parm stack space that needs to be
4391 save, if any. */
4392 int low_to_save = 0, high_to_save = 0;
4393 rtx save_area = 0; /* Place that it is saved. */
4394 #endif
4396 /* Size of the stack reserved for parameter registers. */
4397 int initial_highest_arg_in_use = highest_outgoing_arg_in_use;
4398 char *initial_stack_usage_map = stack_usage_map;
4399 char *stack_usage_map_buf = NULL;
4401 rtx struct_value = targetm.calls.struct_value_rtx (0, 0);
4403 #ifdef REG_PARM_STACK_SPACE
4404 reg_parm_stack_space = REG_PARM_STACK_SPACE ((tree) 0);
4405 #endif
4407 /* By default, library functions cannot throw. */
4408 flags = ECF_NOTHROW;
4410 switch (fn_type)
4412 case LCT_NORMAL:
4413 break;
4414 case LCT_CONST:
4415 flags |= ECF_CONST;
4416 break;
4417 case LCT_PURE:
4418 flags |= ECF_PURE;
4419 break;
4420 case LCT_NORETURN:
4421 flags |= ECF_NORETURN;
4422 break;
4423 case LCT_THROW:
4424 flags &= ~ECF_NOTHROW;
4425 break;
4426 case LCT_RETURNS_TWICE:
4427 flags = ECF_RETURNS_TWICE;
4428 break;
4430 fun = orgfun;
4432 /* Ensure current function's preferred stack boundary is at least
4433 what we need. */
4434 if (crtl->preferred_stack_boundary < PREFERRED_STACK_BOUNDARY)
4435 crtl->preferred_stack_boundary = PREFERRED_STACK_BOUNDARY;
4437 /* If this kind of value comes back in memory,
4438 decide where in memory it should come back. */
4439 if (outmode != VOIDmode)
4441 tfom = lang_hooks.types.type_for_mode (outmode, 0);
4442 if (aggregate_value_p (tfom, 0))
4444 #ifdef PCC_STATIC_STRUCT_RETURN
4445 rtx pointer_reg
4446 = hard_function_value (build_pointer_type (tfom), 0, 0, 0);
4447 mem_value = gen_rtx_MEM (outmode, pointer_reg);
4448 pcc_struct_value = 1;
4449 if (value == 0)
4450 value = gen_reg_rtx (outmode);
4451 #else /* not PCC_STATIC_STRUCT_RETURN */
4452 struct_value_size = GET_MODE_SIZE (outmode);
4453 if (value != 0 && MEM_P (value))
4454 mem_value = value;
4455 else
4456 mem_value = assign_temp (tfom, 1, 1);
4457 #endif
4458 /* This call returns a big structure. */
4459 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
4462 else
4463 tfom = void_type_node;
4465 /* ??? Unfinished: must pass the memory address as an argument. */
4467 /* Copy all the libcall-arguments out of the varargs data
4468 and into a vector ARGVEC.
4470 Compute how to pass each argument. We only support a very small subset
4471 of the full argument passing conventions to limit complexity here since
4472 library functions shouldn't have many args. */
4474 argvec = XALLOCAVEC (struct arg, nargs + 1);
4475 memset (argvec, 0, (nargs + 1) * sizeof (struct arg));
4477 #ifdef INIT_CUMULATIVE_LIBCALL_ARGS
4478 INIT_CUMULATIVE_LIBCALL_ARGS (args_so_far_v, outmode, fun);
4479 #else
4480 INIT_CUMULATIVE_ARGS (args_so_far_v, NULL_TREE, fun, 0, nargs);
4481 #endif
4482 args_so_far = pack_cumulative_args (&args_so_far_v);
4484 args_size.constant = 0;
4485 args_size.var = 0;
4487 count = 0;
4489 push_temp_slots ();
4491 /* If there's a structure value address to be passed,
4492 either pass it in the special place, or pass it as an extra argument. */
4493 if (mem_value && struct_value == 0 && ! pcc_struct_value)
4495 rtx addr = XEXP (mem_value, 0);
4497 nargs++;
4499 /* Make sure it is a reasonable operand for a move or push insn. */
4500 if (!REG_P (addr) && !MEM_P (addr)
4501 && !(CONSTANT_P (addr)
4502 && targetm.legitimate_constant_p (Pmode, addr)))
4503 addr = force_operand (addr, NULL_RTX);
4505 argvec[count].value = addr;
4506 argvec[count].mode = Pmode;
4507 argvec[count].partial = 0;
4509 argvec[count].reg = targetm.calls.function_arg (args_so_far,
4510 Pmode, NULL_TREE, true);
4511 gcc_assert (targetm.calls.arg_partial_bytes (args_so_far, Pmode,
4512 NULL_TREE, 1) == 0);
4514 locate_and_pad_parm (Pmode, NULL_TREE,
4515 #ifdef STACK_PARMS_IN_REG_PARM_AREA
4517 #else
4518 argvec[count].reg != 0,
4519 #endif
4520 reg_parm_stack_space, 0,
4521 NULL_TREE, &args_size, &argvec[count].locate);
4523 if (argvec[count].reg == 0 || argvec[count].partial != 0
4524 || reg_parm_stack_space > 0)
4525 args_size.constant += argvec[count].locate.size.constant;
4527 targetm.calls.function_arg_advance (args_so_far, Pmode, (tree) 0, true);
4529 count++;
4532 for (unsigned int i = 0; count < nargs; i++, count++)
4534 rtx val = args[i].first;
4535 machine_mode mode = args[i].second;
4536 int unsigned_p = 0;
4538 /* We cannot convert the arg value to the mode the library wants here;
4539 must do it earlier where we know the signedness of the arg. */
4540 gcc_assert (mode != BLKmode
4541 && (GET_MODE (val) == mode || GET_MODE (val) == VOIDmode));
4543 /* Make sure it is a reasonable operand for a move or push insn. */
4544 if (!REG_P (val) && !MEM_P (val)
4545 && !(CONSTANT_P (val) && targetm.legitimate_constant_p (mode, val)))
4546 val = force_operand (val, NULL_RTX);
4548 if (pass_by_reference (&args_so_far_v, mode, NULL_TREE, 1))
4550 rtx slot;
4551 int must_copy
4552 = !reference_callee_copied (&args_so_far_v, mode, NULL_TREE, 1);
4554 /* If this was a CONST function, it is now PURE since it now
4555 reads memory. */
4556 if (flags & ECF_CONST)
4558 flags &= ~ECF_CONST;
4559 flags |= ECF_PURE;
4562 if (MEM_P (val) && !must_copy)
4564 tree val_expr = MEM_EXPR (val);
4565 if (val_expr)
4566 mark_addressable (val_expr);
4567 slot = val;
4569 else
4571 slot = assign_temp (lang_hooks.types.type_for_mode (mode, 0),
4572 1, 1);
4573 emit_move_insn (slot, val);
4576 call_fusage = gen_rtx_EXPR_LIST (VOIDmode,
4577 gen_rtx_USE (VOIDmode, slot),
4578 call_fusage);
4579 if (must_copy)
4580 call_fusage = gen_rtx_EXPR_LIST (VOIDmode,
4581 gen_rtx_CLOBBER (VOIDmode,
4582 slot),
4583 call_fusage);
4585 mode = Pmode;
4586 val = force_operand (XEXP (slot, 0), NULL_RTX);
4589 mode = promote_function_mode (NULL_TREE, mode, &unsigned_p, NULL_TREE, 0);
4590 argvec[count].mode = mode;
4591 argvec[count].value = convert_modes (mode, GET_MODE (val), val, unsigned_p);
4592 argvec[count].reg = targetm.calls.function_arg (args_so_far, mode,
4593 NULL_TREE, true);
4595 argvec[count].partial
4596 = targetm.calls.arg_partial_bytes (args_so_far, mode, NULL_TREE, 1);
4598 if (argvec[count].reg == 0
4599 || argvec[count].partial != 0
4600 || reg_parm_stack_space > 0)
4602 locate_and_pad_parm (mode, NULL_TREE,
4603 #ifdef STACK_PARMS_IN_REG_PARM_AREA
4605 #else
4606 argvec[count].reg != 0,
4607 #endif
4608 reg_parm_stack_space, argvec[count].partial,
4609 NULL_TREE, &args_size, &argvec[count].locate);
4610 args_size.constant += argvec[count].locate.size.constant;
4611 gcc_assert (!argvec[count].locate.size.var);
4613 #ifdef BLOCK_REG_PADDING
4614 else
4615 /* The argument is passed entirely in registers. See at which
4616 end it should be padded. */
4617 argvec[count].locate.where_pad =
4618 BLOCK_REG_PADDING (mode, NULL_TREE,
4619 GET_MODE_SIZE (mode) <= UNITS_PER_WORD);
4620 #endif
4622 targetm.calls.function_arg_advance (args_so_far, mode, (tree) 0, true);
4625 /* If this machine requires an external definition for library
4626 functions, write one out. */
4627 assemble_external_libcall (fun);
4629 original_args_size = args_size;
4630 args_size.constant = (((args_size.constant
4631 + stack_pointer_delta
4632 + STACK_BYTES - 1)
4633 / STACK_BYTES
4634 * STACK_BYTES)
4635 - stack_pointer_delta);
4637 args_size.constant = MAX (args_size.constant,
4638 reg_parm_stack_space);
4640 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
4641 args_size.constant -= reg_parm_stack_space;
4643 if (args_size.constant > crtl->outgoing_args_size)
4644 crtl->outgoing_args_size = args_size.constant;
4646 if (flag_stack_usage_info && !ACCUMULATE_OUTGOING_ARGS)
4648 int pushed = args_size.constant + pending_stack_adjust;
4649 if (pushed > current_function_pushed_stack_size)
4650 current_function_pushed_stack_size = pushed;
4653 if (ACCUMULATE_OUTGOING_ARGS)
4655 /* Since the stack pointer will never be pushed, it is possible for
4656 the evaluation of a parm to clobber something we have already
4657 written to the stack. Since most function calls on RISC machines
4658 do not use the stack, this is uncommon, but must work correctly.
4660 Therefore, we save any area of the stack that was already written
4661 and that we are using. Here we set up to do this by making a new
4662 stack usage map from the old one.
4664 Another approach might be to try to reorder the argument
4665 evaluations to avoid this conflicting stack usage. */
4667 needed = args_size.constant;
4669 /* Since we will be writing into the entire argument area, the
4670 map must be allocated for its entire size, not just the part that
4671 is the responsibility of the caller. */
4672 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
4673 needed += reg_parm_stack_space;
4675 if (ARGS_GROW_DOWNWARD)
4676 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use,
4677 needed + 1);
4678 else
4679 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use, needed);
4681 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use);
4682 stack_usage_map = stack_usage_map_buf;
4684 if (initial_highest_arg_in_use)
4685 memcpy (stack_usage_map, initial_stack_usage_map,
4686 initial_highest_arg_in_use);
4688 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use)
4689 memset (&stack_usage_map[initial_highest_arg_in_use], 0,
4690 highest_outgoing_arg_in_use - initial_highest_arg_in_use);
4691 needed = 0;
4693 /* We must be careful to use virtual regs before they're instantiated,
4694 and real regs afterwards. Loop optimization, for example, can create
4695 new libcalls after we've instantiated the virtual regs, and if we
4696 use virtuals anyway, they won't match the rtl patterns. */
4698 if (virtuals_instantiated)
4699 argblock = plus_constant (Pmode, stack_pointer_rtx,
4700 STACK_POINTER_OFFSET);
4701 else
4702 argblock = virtual_outgoing_args_rtx;
4704 else
4706 if (!PUSH_ARGS)
4707 argblock = push_block (GEN_INT (args_size.constant), 0, 0);
4710 /* We push args individually in reverse order, perform stack alignment
4711 before the first push (the last arg). */
4712 if (argblock == 0)
4713 anti_adjust_stack (GEN_INT (args_size.constant
4714 - original_args_size.constant));
4716 argnum = nargs - 1;
4718 #ifdef REG_PARM_STACK_SPACE
4719 if (ACCUMULATE_OUTGOING_ARGS)
4721 /* The argument list is the property of the called routine and it
4722 may clobber it. If the fixed area has been used for previous
4723 parameters, we must save and restore it. */
4724 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock,
4725 &low_to_save, &high_to_save);
4727 #endif
4729 /* When expanding a normal call, args are stored in push order,
4730 which is the reverse of what we have here. */
4731 bool any_regs = false;
4732 for (int i = nargs; i-- > 0; )
4733 if (argvec[i].reg != NULL_RTX)
4735 targetm.calls.call_args (argvec[i].reg, NULL_TREE);
4736 any_regs = true;
4738 if (!any_regs)
4739 targetm.calls.call_args (pc_rtx, NULL_TREE);
4741 /* Push the args that need to be pushed. */
4743 have_push_fusage = false;
4745 /* ARGNUM indexes the ARGVEC array in the order in which the arguments
4746 are to be pushed. */
4747 for (count = 0; count < nargs; count++, argnum--)
4749 machine_mode mode = argvec[argnum].mode;
4750 rtx val = argvec[argnum].value;
4751 rtx reg = argvec[argnum].reg;
4752 int partial = argvec[argnum].partial;
4753 unsigned int parm_align = argvec[argnum].locate.boundary;
4754 int lower_bound = 0, upper_bound = 0, i;
4756 if (! (reg != 0 && partial == 0))
4758 rtx use;
4760 if (ACCUMULATE_OUTGOING_ARGS)
4762 /* If this is being stored into a pre-allocated, fixed-size,
4763 stack area, save any previous data at that location. */
4765 if (ARGS_GROW_DOWNWARD)
4767 /* stack_slot is negative, but we want to index stack_usage_map
4768 with positive values. */
4769 upper_bound = -argvec[argnum].locate.slot_offset.constant + 1;
4770 lower_bound = upper_bound - argvec[argnum].locate.size.constant;
4772 else
4774 lower_bound = argvec[argnum].locate.slot_offset.constant;
4775 upper_bound = lower_bound + argvec[argnum].locate.size.constant;
4778 i = lower_bound;
4779 /* Don't worry about things in the fixed argument area;
4780 it has already been saved. */
4781 if (i < reg_parm_stack_space)
4782 i = reg_parm_stack_space;
4783 while (i < upper_bound && stack_usage_map[i] == 0)
4784 i++;
4786 if (i < upper_bound)
4788 /* We need to make a save area. */
4789 unsigned int size
4790 = argvec[argnum].locate.size.constant * BITS_PER_UNIT;
4791 machine_mode save_mode
4792 = int_mode_for_size (size, 1).else_blk ();
4793 rtx adr
4794 = plus_constant (Pmode, argblock,
4795 argvec[argnum].locate.offset.constant);
4796 rtx stack_area
4797 = gen_rtx_MEM (save_mode, memory_address (save_mode, adr));
4799 if (save_mode == BLKmode)
4801 argvec[argnum].save_area
4802 = assign_stack_temp (BLKmode,
4803 argvec[argnum].locate.size.constant
4806 emit_block_move (validize_mem
4807 (copy_rtx (argvec[argnum].save_area)),
4808 stack_area,
4809 GEN_INT (argvec[argnum].locate.size.constant),
4810 BLOCK_OP_CALL_PARM);
4812 else
4814 argvec[argnum].save_area = gen_reg_rtx (save_mode);
4816 emit_move_insn (argvec[argnum].save_area, stack_area);
4821 emit_push_insn (val, mode, NULL_TREE, NULL_RTX, parm_align,
4822 partial, reg, 0, argblock,
4823 GEN_INT (argvec[argnum].locate.offset.constant),
4824 reg_parm_stack_space,
4825 ARGS_SIZE_RTX (argvec[argnum].locate.alignment_pad), false);
4827 /* Now mark the segment we just used. */
4828 if (ACCUMULATE_OUTGOING_ARGS)
4829 for (i = lower_bound; i < upper_bound; i++)
4830 stack_usage_map[i] = 1;
4832 NO_DEFER_POP;
4834 /* Indicate argument access so that alias.c knows that these
4835 values are live. */
4836 if (argblock)
4837 use = plus_constant (Pmode, argblock,
4838 argvec[argnum].locate.offset.constant);
4839 else if (have_push_fusage)
4840 continue;
4841 else
4843 /* When arguments are pushed, trying to tell alias.c where
4844 exactly this argument is won't work, because the
4845 auto-increment causes confusion. So we merely indicate
4846 that we access something with a known mode somewhere on
4847 the stack. */
4848 use = gen_rtx_PLUS (Pmode, stack_pointer_rtx,
4849 gen_rtx_SCRATCH (Pmode));
4850 have_push_fusage = true;
4852 use = gen_rtx_MEM (argvec[argnum].mode, use);
4853 use = gen_rtx_USE (VOIDmode, use);
4854 call_fusage = gen_rtx_EXPR_LIST (VOIDmode, use, call_fusage);
4858 argnum = nargs - 1;
4860 fun = prepare_call_address (NULL, fun, NULL, &call_fusage, 0, 0);
4862 /* Now load any reg parms into their regs. */
4864 /* ARGNUM indexes the ARGVEC array in the order in which the arguments
4865 are to be pushed. */
4866 for (count = 0; count < nargs; count++, argnum--)
4868 machine_mode mode = argvec[argnum].mode;
4869 rtx val = argvec[argnum].value;
4870 rtx reg = argvec[argnum].reg;
4871 int partial = argvec[argnum].partial;
4872 #ifdef BLOCK_REG_PADDING
4873 int size = 0;
4874 #endif
4876 /* Handle calls that pass values in multiple non-contiguous
4877 locations. The PA64 has examples of this for library calls. */
4878 if (reg != 0 && GET_CODE (reg) == PARALLEL)
4879 emit_group_load (reg, val, NULL_TREE, GET_MODE_SIZE (mode));
4880 else if (reg != 0 && partial == 0)
4882 emit_move_insn (reg, val);
4883 #ifdef BLOCK_REG_PADDING
4884 size = GET_MODE_SIZE (argvec[argnum].mode);
4886 /* Copied from load_register_parameters. */
4888 /* Handle case where we have a value that needs shifting
4889 up to the msb. eg. a QImode value and we're padding
4890 upward on a BYTES_BIG_ENDIAN machine. */
4891 if (size < UNITS_PER_WORD
4892 && (argvec[argnum].locate.where_pad
4893 == (BYTES_BIG_ENDIAN ? PAD_UPWARD : PAD_DOWNWARD)))
4895 rtx x;
4896 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
4898 /* Assigning REG here rather than a temp makes CALL_FUSAGE
4899 report the whole reg as used. Strictly speaking, the
4900 call only uses SIZE bytes at the msb end, but it doesn't
4901 seem worth generating rtl to say that. */
4902 reg = gen_rtx_REG (word_mode, REGNO (reg));
4903 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1);
4904 if (x != reg)
4905 emit_move_insn (reg, x);
4907 #endif
4910 NO_DEFER_POP;
4913 /* Any regs containing parms remain in use through the call. */
4914 for (count = 0; count < nargs; count++)
4916 rtx reg = argvec[count].reg;
4917 if (reg != 0 && GET_CODE (reg) == PARALLEL)
4918 use_group_regs (&call_fusage, reg);
4919 else if (reg != 0)
4921 int partial = argvec[count].partial;
4922 if (partial)
4924 int nregs;
4925 gcc_assert (partial % UNITS_PER_WORD == 0);
4926 nregs = partial / UNITS_PER_WORD;
4927 use_regs (&call_fusage, REGNO (reg), nregs);
4929 else
4930 use_reg (&call_fusage, reg);
4934 /* Pass the function the address in which to return a structure value. */
4935 if (mem_value != 0 && struct_value != 0 && ! pcc_struct_value)
4937 emit_move_insn (struct_value,
4938 force_reg (Pmode,
4939 force_operand (XEXP (mem_value, 0),
4940 NULL_RTX)));
4941 if (REG_P (struct_value))
4942 use_reg (&call_fusage, struct_value);
4945 /* Don't allow popping to be deferred, since then
4946 cse'ing of library calls could delete a call and leave the pop. */
4947 NO_DEFER_POP;
4948 valreg = (mem_value == 0 && outmode != VOIDmode
4949 ? hard_libcall_value (outmode, orgfun) : NULL_RTX);
4951 /* Stack must be properly aligned now. */
4952 gcc_assert (!(stack_pointer_delta
4953 & (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT - 1)));
4955 before_call = get_last_insn ();
4957 /* We pass the old value of inhibit_defer_pop + 1 to emit_call_1, which
4958 will set inhibit_defer_pop to that value. */
4959 /* The return type is needed to decide how many bytes the function pops.
4960 Signedness plays no role in that, so for simplicity, we pretend it's
4961 always signed. We also assume that the list of arguments passed has
4962 no impact, so we pretend it is unknown. */
4964 emit_call_1 (fun, NULL,
4965 get_identifier (XSTR (orgfun, 0)),
4966 build_function_type (tfom, NULL_TREE),
4967 original_args_size.constant, args_size.constant,
4968 struct_value_size,
4969 targetm.calls.function_arg (args_so_far,
4970 VOIDmode, void_type_node, true),
4971 valreg,
4972 old_inhibit_defer_pop + 1, call_fusage, flags, args_so_far);
4974 if (flag_ipa_ra)
4976 rtx datum = orgfun;
4977 gcc_assert (GET_CODE (datum) == SYMBOL_REF);
4978 rtx_call_insn *last = last_call_insn ();
4979 add_reg_note (last, REG_CALL_DECL, datum);
4982 /* Right-shift returned value if necessary. */
4983 if (!pcc_struct_value
4984 && TYPE_MODE (tfom) != BLKmode
4985 && targetm.calls.return_in_msb (tfom))
4987 shift_return_value (TYPE_MODE (tfom), false, valreg);
4988 valreg = gen_rtx_REG (TYPE_MODE (tfom), REGNO (valreg));
4991 targetm.calls.end_call_args ();
4993 /* For calls to `setjmp', etc., inform function.c:setjmp_warnings
4994 that it should complain if nonvolatile values are live. For
4995 functions that cannot return, inform flow that control does not
4996 fall through. */
4997 if (flags & ECF_NORETURN)
4999 /* The barrier note must be emitted
5000 immediately after the CALL_INSN. Some ports emit more than
5001 just a CALL_INSN above, so we must search for it here. */
5002 rtx_insn *last = get_last_insn ();
5003 while (!CALL_P (last))
5005 last = PREV_INSN (last);
5006 /* There was no CALL_INSN? */
5007 gcc_assert (last != before_call);
5010 emit_barrier_after (last);
5013 /* Consider that "regular" libcalls, i.e. all of them except for LCT_THROW
5014 and LCT_RETURNS_TWICE, cannot perform non-local gotos. */
5015 if (flags & ECF_NOTHROW)
5017 rtx_insn *last = get_last_insn ();
5018 while (!CALL_P (last))
5020 last = PREV_INSN (last);
5021 /* There was no CALL_INSN? */
5022 gcc_assert (last != before_call);
5025 make_reg_eh_region_note_nothrow_nononlocal (last);
5028 /* Now restore inhibit_defer_pop to its actual original value. */
5029 OK_DEFER_POP;
5031 pop_temp_slots ();
5033 /* Copy the value to the right place. */
5034 if (outmode != VOIDmode && retval)
5036 if (mem_value)
5038 if (value == 0)
5039 value = mem_value;
5040 if (value != mem_value)
5041 emit_move_insn (value, mem_value);
5043 else if (GET_CODE (valreg) == PARALLEL)
5045 if (value == 0)
5046 value = gen_reg_rtx (outmode);
5047 emit_group_store (value, valreg, NULL_TREE, GET_MODE_SIZE (outmode));
5049 else
5051 /* Convert to the proper mode if a promotion has been active. */
5052 if (GET_MODE (valreg) != outmode)
5054 int unsignedp = TYPE_UNSIGNED (tfom);
5056 gcc_assert (promote_function_mode (tfom, outmode, &unsignedp,
5057 fndecl ? TREE_TYPE (fndecl) : fntype, 1)
5058 == GET_MODE (valreg));
5059 valreg = convert_modes (outmode, GET_MODE (valreg), valreg, 0);
5062 if (value != 0)
5063 emit_move_insn (value, valreg);
5064 else
5065 value = valreg;
5069 if (ACCUMULATE_OUTGOING_ARGS)
5071 #ifdef REG_PARM_STACK_SPACE
5072 if (save_area)
5073 restore_fixed_argument_area (save_area, argblock,
5074 high_to_save, low_to_save);
5075 #endif
5077 /* If we saved any argument areas, restore them. */
5078 for (count = 0; count < nargs; count++)
5079 if (argvec[count].save_area)
5081 machine_mode save_mode = GET_MODE (argvec[count].save_area);
5082 rtx adr = plus_constant (Pmode, argblock,
5083 argvec[count].locate.offset.constant);
5084 rtx stack_area = gen_rtx_MEM (save_mode,
5085 memory_address (save_mode, adr));
5087 if (save_mode == BLKmode)
5088 emit_block_move (stack_area,
5089 validize_mem
5090 (copy_rtx (argvec[count].save_area)),
5091 GEN_INT (argvec[count].locate.size.constant),
5092 BLOCK_OP_CALL_PARM);
5093 else
5094 emit_move_insn (stack_area, argvec[count].save_area);
5097 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
5098 stack_usage_map = initial_stack_usage_map;
5101 free (stack_usage_map_buf);
5103 return value;
5108 /* Store pointer bounds argument ARG into Bounds Table entry
5109 associated with PARM. */
5110 static void
5111 store_bounds (struct arg_data *arg, struct arg_data *parm)
5113 rtx slot = NULL, ptr = NULL, addr = NULL;
5115 /* We may pass bounds not associated with any pointer. */
5116 if (!parm)
5118 gcc_assert (arg->special_slot);
5119 slot = arg->special_slot;
5120 ptr = const0_rtx;
5122 /* Find pointer associated with bounds and where it is
5123 passed. */
5124 else
5126 if (!parm->reg)
5128 gcc_assert (!arg->special_slot);
5130 addr = adjust_address (parm->stack, Pmode, arg->pointer_offset);
5132 else if (REG_P (parm->reg))
5134 gcc_assert (arg->special_slot);
5135 slot = arg->special_slot;
5137 if (MEM_P (parm->value))
5138 addr = adjust_address (parm->value, Pmode, arg->pointer_offset);
5139 else if (REG_P (parm->value))
5140 ptr = gen_rtx_SUBREG (Pmode, parm->value, arg->pointer_offset);
5141 else
5143 gcc_assert (!arg->pointer_offset);
5144 ptr = parm->value;
5147 else
5149 gcc_assert (GET_CODE (parm->reg) == PARALLEL);
5151 gcc_assert (arg->special_slot);
5152 slot = arg->special_slot;
5154 if (parm->parallel_value)
5155 ptr = chkp_get_value_with_offs (parm->parallel_value,
5156 GEN_INT (arg->pointer_offset));
5157 else
5158 gcc_unreachable ();
5162 /* Expand bounds. */
5163 if (!arg->value)
5164 arg->value = expand_normal (arg->tree_value);
5166 targetm.calls.store_bounds_for_arg (ptr, addr, arg->value, slot);
5169 /* Store a single argument for a function call
5170 into the register or memory area where it must be passed.
5171 *ARG describes the argument value and where to pass it.
5173 ARGBLOCK is the address of the stack-block for all the arguments,
5174 or 0 on a machine where arguments are pushed individually.
5176 MAY_BE_ALLOCA nonzero says this could be a call to `alloca'
5177 so must be careful about how the stack is used.
5179 VARIABLE_SIZE nonzero says that this was a variable-sized outgoing
5180 argument stack. This is used if ACCUMULATE_OUTGOING_ARGS to indicate
5181 that we need not worry about saving and restoring the stack.
5183 FNDECL is the declaration of the function we are calling.
5185 Return nonzero if this arg should cause sibcall failure,
5186 zero otherwise. */
5188 static int
5189 store_one_arg (struct arg_data *arg, rtx argblock, int flags,
5190 int variable_size ATTRIBUTE_UNUSED, int reg_parm_stack_space)
5192 tree pval = arg->tree_value;
5193 rtx reg = 0;
5194 int partial = 0;
5195 int used = 0;
5196 int i, lower_bound = 0, upper_bound = 0;
5197 int sibcall_failure = 0;
5199 if (TREE_CODE (pval) == ERROR_MARK)
5200 return 1;
5202 /* Push a new temporary level for any temporaries we make for
5203 this argument. */
5204 push_temp_slots ();
5206 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL))
5208 /* If this is being stored into a pre-allocated, fixed-size, stack area,
5209 save any previous data at that location. */
5210 if (argblock && ! variable_size && arg->stack)
5212 if (ARGS_GROW_DOWNWARD)
5214 /* stack_slot is negative, but we want to index stack_usage_map
5215 with positive values. */
5216 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS)
5217 upper_bound = -INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1)) + 1;
5218 else
5219 upper_bound = 0;
5221 lower_bound = upper_bound - arg->locate.size.constant;
5223 else
5225 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS)
5226 lower_bound = INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1));
5227 else
5228 lower_bound = 0;
5230 upper_bound = lower_bound + arg->locate.size.constant;
5233 i = lower_bound;
5234 /* Don't worry about things in the fixed argument area;
5235 it has already been saved. */
5236 if (i < reg_parm_stack_space)
5237 i = reg_parm_stack_space;
5238 while (i < upper_bound && stack_usage_map[i] == 0)
5239 i++;
5241 if (i < upper_bound)
5243 /* We need to make a save area. */
5244 unsigned int size = arg->locate.size.constant * BITS_PER_UNIT;
5245 machine_mode save_mode
5246 = int_mode_for_size (size, 1).else_blk ();
5247 rtx adr = memory_address (save_mode, XEXP (arg->stack_slot, 0));
5248 rtx stack_area = gen_rtx_MEM (save_mode, adr);
5250 if (save_mode == BLKmode)
5252 arg->save_area
5253 = assign_temp (TREE_TYPE (arg->tree_value), 1, 1);
5254 preserve_temp_slots (arg->save_area);
5255 emit_block_move (validize_mem (copy_rtx (arg->save_area)),
5256 stack_area,
5257 GEN_INT (arg->locate.size.constant),
5258 BLOCK_OP_CALL_PARM);
5260 else
5262 arg->save_area = gen_reg_rtx (save_mode);
5263 emit_move_insn (arg->save_area, stack_area);
5269 /* If this isn't going to be placed on both the stack and in registers,
5270 set up the register and number of words. */
5271 if (! arg->pass_on_stack)
5273 if (flags & ECF_SIBCALL)
5274 reg = arg->tail_call_reg;
5275 else
5276 reg = arg->reg;
5277 partial = arg->partial;
5280 /* Being passed entirely in a register. We shouldn't be called in
5281 this case. */
5282 gcc_assert (reg == 0 || partial != 0);
5284 /* If this arg needs special alignment, don't load the registers
5285 here. */
5286 if (arg->n_aligned_regs != 0)
5287 reg = 0;
5289 /* If this is being passed partially in a register, we can't evaluate
5290 it directly into its stack slot. Otherwise, we can. */
5291 if (arg->value == 0)
5293 /* stack_arg_under_construction is nonzero if a function argument is
5294 being evaluated directly into the outgoing argument list and
5295 expand_call must take special action to preserve the argument list
5296 if it is called recursively.
5298 For scalar function arguments stack_usage_map is sufficient to
5299 determine which stack slots must be saved and restored. Scalar
5300 arguments in general have pass_on_stack == 0.
5302 If this argument is initialized by a function which takes the
5303 address of the argument (a C++ constructor or a C function
5304 returning a BLKmode structure), then stack_usage_map is
5305 insufficient and expand_call must push the stack around the
5306 function call. Such arguments have pass_on_stack == 1.
5308 Note that it is always safe to set stack_arg_under_construction,
5309 but this generates suboptimal code if set when not needed. */
5311 if (arg->pass_on_stack)
5312 stack_arg_under_construction++;
5314 arg->value = expand_expr (pval,
5315 (partial
5316 || TYPE_MODE (TREE_TYPE (pval)) != arg->mode)
5317 ? NULL_RTX : arg->stack,
5318 VOIDmode, EXPAND_STACK_PARM);
5320 /* If we are promoting object (or for any other reason) the mode
5321 doesn't agree, convert the mode. */
5323 if (arg->mode != TYPE_MODE (TREE_TYPE (pval)))
5324 arg->value = convert_modes (arg->mode, TYPE_MODE (TREE_TYPE (pval)),
5325 arg->value, arg->unsignedp);
5327 if (arg->pass_on_stack)
5328 stack_arg_under_construction--;
5331 /* Check for overlap with already clobbered argument area. */
5332 if ((flags & ECF_SIBCALL)
5333 && MEM_P (arg->value)
5334 && mem_overlaps_already_clobbered_arg_p (XEXP (arg->value, 0),
5335 arg->locate.size.constant))
5336 sibcall_failure = 1;
5338 /* Don't allow anything left on stack from computation
5339 of argument to alloca. */
5340 if (flags & ECF_MAY_BE_ALLOCA)
5341 do_pending_stack_adjust ();
5343 if (arg->value == arg->stack)
5344 /* If the value is already in the stack slot, we are done. */
5346 else if (arg->mode != BLKmode)
5348 int size;
5349 unsigned int parm_align;
5351 /* Argument is a scalar, not entirely passed in registers.
5352 (If part is passed in registers, arg->partial says how much
5353 and emit_push_insn will take care of putting it there.)
5355 Push it, and if its size is less than the
5356 amount of space allocated to it,
5357 also bump stack pointer by the additional space.
5358 Note that in C the default argument promotions
5359 will prevent such mismatches. */
5361 size = GET_MODE_SIZE (arg->mode);
5362 /* Compute how much space the push instruction will push.
5363 On many machines, pushing a byte will advance the stack
5364 pointer by a halfword. */
5365 #ifdef PUSH_ROUNDING
5366 size = PUSH_ROUNDING (size);
5367 #endif
5368 used = size;
5370 /* Compute how much space the argument should get:
5371 round up to a multiple of the alignment for arguments. */
5372 if (targetm.calls.function_arg_padding (arg->mode, TREE_TYPE (pval))
5373 != PAD_NONE)
5374 used = (((size + PARM_BOUNDARY / BITS_PER_UNIT - 1)
5375 / (PARM_BOUNDARY / BITS_PER_UNIT))
5376 * (PARM_BOUNDARY / BITS_PER_UNIT));
5378 /* Compute the alignment of the pushed argument. */
5379 parm_align = arg->locate.boundary;
5380 if (targetm.calls.function_arg_padding (arg->mode, TREE_TYPE (pval))
5381 == PAD_DOWNWARD)
5383 int pad = used - size;
5384 if (pad)
5386 unsigned int pad_align = least_bit_hwi (pad) * BITS_PER_UNIT;
5387 parm_align = MIN (parm_align, pad_align);
5391 /* This isn't already where we want it on the stack, so put it there.
5392 This can either be done with push or copy insns. */
5393 if (!emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), NULL_RTX,
5394 parm_align, partial, reg, used - size, argblock,
5395 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space,
5396 ARGS_SIZE_RTX (arg->locate.alignment_pad), true))
5397 sibcall_failure = 1;
5399 /* Unless this is a partially-in-register argument, the argument is now
5400 in the stack. */
5401 if (partial == 0)
5402 arg->value = arg->stack;
5404 else
5406 /* BLKmode, at least partly to be pushed. */
5408 unsigned int parm_align;
5409 int excess;
5410 rtx size_rtx;
5412 /* Pushing a nonscalar.
5413 If part is passed in registers, PARTIAL says how much
5414 and emit_push_insn will take care of putting it there. */
5416 /* Round its size up to a multiple
5417 of the allocation unit for arguments. */
5419 if (arg->locate.size.var != 0)
5421 excess = 0;
5422 size_rtx = ARGS_SIZE_RTX (arg->locate.size);
5424 else
5426 /* PUSH_ROUNDING has no effect on us, because emit_push_insn
5427 for BLKmode is careful to avoid it. */
5428 excess = (arg->locate.size.constant
5429 - int_size_in_bytes (TREE_TYPE (pval))
5430 + partial);
5431 size_rtx = expand_expr (size_in_bytes (TREE_TYPE (pval)),
5432 NULL_RTX, TYPE_MODE (sizetype),
5433 EXPAND_NORMAL);
5436 parm_align = arg->locate.boundary;
5438 /* When an argument is padded down, the block is aligned to
5439 PARM_BOUNDARY, but the actual argument isn't. */
5440 if (targetm.calls.function_arg_padding (arg->mode, TREE_TYPE (pval))
5441 == PAD_DOWNWARD)
5443 if (arg->locate.size.var)
5444 parm_align = BITS_PER_UNIT;
5445 else if (excess)
5447 unsigned int excess_align = least_bit_hwi (excess) * BITS_PER_UNIT;
5448 parm_align = MIN (parm_align, excess_align);
5452 if ((flags & ECF_SIBCALL) && MEM_P (arg->value))
5454 /* emit_push_insn might not work properly if arg->value and
5455 argblock + arg->locate.offset areas overlap. */
5456 rtx x = arg->value;
5457 int i = 0;
5459 if (XEXP (x, 0) == crtl->args.internal_arg_pointer
5460 || (GET_CODE (XEXP (x, 0)) == PLUS
5461 && XEXP (XEXP (x, 0), 0) ==
5462 crtl->args.internal_arg_pointer
5463 && CONST_INT_P (XEXP (XEXP (x, 0), 1))))
5465 if (XEXP (x, 0) != crtl->args.internal_arg_pointer)
5466 i = INTVAL (XEXP (XEXP (x, 0), 1));
5468 /* arg.locate doesn't contain the pretend_args_size offset,
5469 it's part of argblock. Ensure we don't count it in I. */
5470 if (STACK_GROWS_DOWNWARD)
5471 i -= crtl->args.pretend_args_size;
5472 else
5473 i += crtl->args.pretend_args_size;
5475 /* expand_call should ensure this. */
5476 gcc_assert (!arg->locate.offset.var
5477 && arg->locate.size.var == 0
5478 && CONST_INT_P (size_rtx));
5480 if (arg->locate.offset.constant > i)
5482 if (arg->locate.offset.constant < i + INTVAL (size_rtx))
5483 sibcall_failure = 1;
5485 else if (arg->locate.offset.constant < i)
5487 /* Use arg->locate.size.constant instead of size_rtx
5488 because we only care about the part of the argument
5489 on the stack. */
5490 if (i < (arg->locate.offset.constant
5491 + arg->locate.size.constant))
5492 sibcall_failure = 1;
5494 else
5496 /* Even though they appear to be at the same location,
5497 if part of the outgoing argument is in registers,
5498 they aren't really at the same location. Check for
5499 this by making sure that the incoming size is the
5500 same as the outgoing size. */
5501 if (arg->locate.size.constant != INTVAL (size_rtx))
5502 sibcall_failure = 1;
5507 emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), size_rtx,
5508 parm_align, partial, reg, excess, argblock,
5509 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space,
5510 ARGS_SIZE_RTX (arg->locate.alignment_pad), false);
5512 /* Unless this is a partially-in-register argument, the argument is now
5513 in the stack.
5515 ??? Unlike the case above, in which we want the actual
5516 address of the data, so that we can load it directly into a
5517 register, here we want the address of the stack slot, so that
5518 it's properly aligned for word-by-word copying or something
5519 like that. It's not clear that this is always correct. */
5520 if (partial == 0)
5521 arg->value = arg->stack_slot;
5524 if (arg->reg && GET_CODE (arg->reg) == PARALLEL)
5526 tree type = TREE_TYPE (arg->tree_value);
5527 arg->parallel_value
5528 = emit_group_load_into_temps (arg->reg, arg->value, type,
5529 int_size_in_bytes (type));
5532 /* Mark all slots this store used. */
5533 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL)
5534 && argblock && ! variable_size && arg->stack)
5535 for (i = lower_bound; i < upper_bound; i++)
5536 stack_usage_map[i] = 1;
5538 /* Once we have pushed something, pops can't safely
5539 be deferred during the rest of the arguments. */
5540 NO_DEFER_POP;
5542 /* Free any temporary slots made in processing this argument. */
5543 pop_temp_slots ();
5545 return sibcall_failure;
5548 /* Nonzero if we do not know how to pass TYPE solely in registers. */
5550 bool
5551 must_pass_in_stack_var_size (machine_mode mode ATTRIBUTE_UNUSED,
5552 const_tree type)
5554 if (!type)
5555 return false;
5557 /* If the type has variable size... */
5558 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
5559 return true;
5561 /* If the type is marked as addressable (it is required
5562 to be constructed into the stack)... */
5563 if (TREE_ADDRESSABLE (type))
5564 return true;
5566 return false;
5569 /* Another version of the TARGET_MUST_PASS_IN_STACK hook. This one
5570 takes trailing padding of a structure into account. */
5571 /* ??? Should be able to merge these two by examining BLOCK_REG_PADDING. */
5573 bool
5574 must_pass_in_stack_var_size_or_pad (machine_mode mode, const_tree type)
5576 if (!type)
5577 return false;
5579 /* If the type has variable size... */
5580 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
5581 return true;
5583 /* If the type is marked as addressable (it is required
5584 to be constructed into the stack)... */
5585 if (TREE_ADDRESSABLE (type))
5586 return true;
5588 /* If the padding and mode of the type is such that a copy into
5589 a register would put it into the wrong part of the register. */
5590 if (mode == BLKmode
5591 && int_size_in_bytes (type) % (PARM_BOUNDARY / BITS_PER_UNIT)
5592 && (targetm.calls.function_arg_padding (mode, type)
5593 == (BYTES_BIG_ENDIAN ? PAD_UPWARD : PAD_DOWNWARD)))
5594 return true;
5596 return false;
5599 /* Tell the garbage collector about GTY markers in this source file. */
5600 #include "gt-calls.h"