PR tree-optimization/78496
[official-gcc.git] / gcc / calls.c
blobc26f157b0b11cf2b991a55c3aa29e508d0498532
1 /* Convert function calls to rtl insns, for GNU C compiler.
2 Copyright (C) 1989-2017 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "target.h"
25 #include "rtl.h"
26 #include "tree.h"
27 #include "gimple.h"
28 #include "predict.h"
29 #include "memmodel.h"
30 #include "tm_p.h"
31 #include "stringpool.h"
32 #include "expmed.h"
33 #include "optabs.h"
34 #include "emit-rtl.h"
35 #include "cgraph.h"
36 #include "diagnostic-core.h"
37 #include "fold-const.h"
38 #include "stor-layout.h"
39 #include "varasm.h"
40 #include "internal-fn.h"
41 #include "dojump.h"
42 #include "explow.h"
43 #include "calls.h"
44 #include "expr.h"
45 #include "output.h"
46 #include "langhooks.h"
47 #include "except.h"
48 #include "dbgcnt.h"
49 #include "rtl-iter.h"
50 #include "tree-chkp.h"
51 #include "tree-vrp.h"
52 #include "tree-ssanames.h"
53 #include "rtl-chkp.h"
54 #include "intl.h"
56 /* Like PREFERRED_STACK_BOUNDARY but in units of bytes, not bits. */
57 #define STACK_BYTES (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT)
59 /* Data structure and subroutines used within expand_call. */
61 struct arg_data
63 /* Tree node for this argument. */
64 tree tree_value;
65 /* Mode for value; TYPE_MODE unless promoted. */
66 machine_mode mode;
67 /* Current RTL value for argument, or 0 if it isn't precomputed. */
68 rtx value;
69 /* Initially-compute RTL value for argument; only for const functions. */
70 rtx initial_value;
71 /* Register to pass this argument in, 0 if passed on stack, or an
72 PARALLEL if the arg is to be copied into multiple non-contiguous
73 registers. */
74 rtx reg;
75 /* Register to pass this argument in when generating tail call sequence.
76 This is not the same register as for normal calls on machines with
77 register windows. */
78 rtx tail_call_reg;
79 /* If REG is a PARALLEL, this is a copy of VALUE pulled into the correct
80 form for emit_group_move. */
81 rtx parallel_value;
82 /* If value is passed in neither reg nor stack, this field holds a number
83 of a special slot to be used. */
84 rtx special_slot;
85 /* For pointer bounds hold an index of parm bounds are bound to. -1 if
86 there is no such pointer. */
87 int pointer_arg;
88 /* If pointer_arg refers a structure, then pointer_offset holds an offset
89 of a pointer in this structure. */
90 int pointer_offset;
91 /* If REG was promoted from the actual mode of the argument expression,
92 indicates whether the promotion is sign- or zero-extended. */
93 int unsignedp;
94 /* Number of bytes to put in registers. 0 means put the whole arg
95 in registers. Also 0 if not passed in registers. */
96 int partial;
97 /* Nonzero if argument must be passed on stack.
98 Note that some arguments may be passed on the stack
99 even though pass_on_stack is zero, just because FUNCTION_ARG says so.
100 pass_on_stack identifies arguments that *cannot* go in registers. */
101 int pass_on_stack;
102 /* Some fields packaged up for locate_and_pad_parm. */
103 struct locate_and_pad_arg_data locate;
104 /* Location on the stack at which parameter should be stored. The store
105 has already been done if STACK == VALUE. */
106 rtx stack;
107 /* Location on the stack of the start of this argument slot. This can
108 differ from STACK if this arg pads downward. This location is known
109 to be aligned to TARGET_FUNCTION_ARG_BOUNDARY. */
110 rtx stack_slot;
111 /* Place that this stack area has been saved, if needed. */
112 rtx save_area;
113 /* If an argument's alignment does not permit direct copying into registers,
114 copy in smaller-sized pieces into pseudos. These are stored in a
115 block pointed to by this field. The next field says how many
116 word-sized pseudos we made. */
117 rtx *aligned_regs;
118 int n_aligned_regs;
121 /* A vector of one char per byte of stack space. A byte if nonzero if
122 the corresponding stack location has been used.
123 This vector is used to prevent a function call within an argument from
124 clobbering any stack already set up. */
125 static char *stack_usage_map;
127 /* Size of STACK_USAGE_MAP. */
128 static int highest_outgoing_arg_in_use;
130 /* A bitmap of virtual-incoming stack space. Bit is set if the corresponding
131 stack location's tail call argument has been already stored into the stack.
132 This bitmap is used to prevent sibling call optimization if function tries
133 to use parent's incoming argument slots when they have been already
134 overwritten with tail call arguments. */
135 static sbitmap stored_args_map;
137 /* stack_arg_under_construction is nonzero when an argument may be
138 initialized with a constructor call (including a C function that
139 returns a BLKmode struct) and expand_call must take special action
140 to make sure the object being constructed does not overlap the
141 argument list for the constructor call. */
142 static int stack_arg_under_construction;
144 static void emit_call_1 (rtx, tree, tree, tree, HOST_WIDE_INT, HOST_WIDE_INT,
145 HOST_WIDE_INT, rtx, rtx, int, rtx, int,
146 cumulative_args_t);
147 static void precompute_register_parameters (int, struct arg_data *, int *);
148 static void store_bounds (struct arg_data *, struct arg_data *);
149 static int store_one_arg (struct arg_data *, rtx, int, int, int);
150 static void store_unaligned_arguments_into_pseudos (struct arg_data *, int);
151 static int finalize_must_preallocate (int, int, struct arg_data *,
152 struct args_size *);
153 static void precompute_arguments (int, struct arg_data *);
154 static int compute_argument_block_size (int, struct args_size *, tree, tree, int);
155 static void initialize_argument_information (int, struct arg_data *,
156 struct args_size *, int,
157 tree, tree,
158 tree, tree, cumulative_args_t, int,
159 rtx *, int *, int *, int *,
160 bool *, bool);
161 static void compute_argument_addresses (struct arg_data *, rtx, int);
162 static rtx rtx_for_function_call (tree, tree);
163 static void load_register_parameters (struct arg_data *, int, rtx *, int,
164 int, int *);
165 static rtx emit_library_call_value_1 (int, rtx, rtx, enum libcall_type,
166 machine_mode, int, va_list);
167 static int special_function_p (const_tree, int);
168 static int check_sibcall_argument_overlap_1 (rtx);
169 static int check_sibcall_argument_overlap (rtx_insn *, struct arg_data *, int);
171 static int combine_pending_stack_adjustment_and_call (int, struct args_size *,
172 unsigned int);
173 static tree split_complex_types (tree);
175 #ifdef REG_PARM_STACK_SPACE
176 static rtx save_fixed_argument_area (int, rtx, int *, int *);
177 static void restore_fixed_argument_area (rtx, rtx, int, int);
178 #endif
180 /* Force FUNEXP into a form suitable for the address of a CALL,
181 and return that as an rtx. Also load the static chain register
182 if FNDECL is a nested function.
184 CALL_FUSAGE points to a variable holding the prospective
185 CALL_INSN_FUNCTION_USAGE information. */
188 prepare_call_address (tree fndecl_or_type, rtx funexp, rtx static_chain_value,
189 rtx *call_fusage, int reg_parm_seen, int flags)
191 /* Make a valid memory address and copy constants through pseudo-regs,
192 but not for a constant address if -fno-function-cse. */
193 if (GET_CODE (funexp) != SYMBOL_REF)
195 /* If it's an indirect call by descriptor, generate code to perform
196 runtime identification of the pointer and load the descriptor. */
197 if ((flags & ECF_BY_DESCRIPTOR) && !flag_trampolines)
199 const int bit_val = targetm.calls.custom_function_descriptors;
200 rtx call_lab = gen_label_rtx ();
202 gcc_assert (fndecl_or_type && TYPE_P (fndecl_or_type));
203 fndecl_or_type
204 = build_decl (UNKNOWN_LOCATION, FUNCTION_DECL, NULL_TREE,
205 fndecl_or_type);
206 DECL_STATIC_CHAIN (fndecl_or_type) = 1;
207 rtx chain = targetm.calls.static_chain (fndecl_or_type, false);
209 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 switch (DECL_FUNCTION_CODE (fndecl))
613 case BUILT_IN_ALLOCA:
614 case BUILT_IN_ALLOCA_WITH_ALIGN:
615 flags |= ECF_MAY_BE_ALLOCA;
616 break;
617 default:
618 break;
621 return flags;
624 /* Similar to special_function_p; return a set of ERF_ flags for the
625 function FNDECL. */
626 static int
627 decl_return_flags (tree fndecl)
629 tree attr;
630 tree type = TREE_TYPE (fndecl);
631 if (!type)
632 return 0;
634 attr = lookup_attribute ("fn spec", TYPE_ATTRIBUTES (type));
635 if (!attr)
636 return 0;
638 attr = TREE_VALUE (TREE_VALUE (attr));
639 if (!attr || TREE_STRING_LENGTH (attr) < 1)
640 return 0;
642 switch (TREE_STRING_POINTER (attr)[0])
644 case '1':
645 case '2':
646 case '3':
647 case '4':
648 return ERF_RETURNS_ARG | (TREE_STRING_POINTER (attr)[0] - '1');
650 case 'm':
651 return ERF_NOALIAS;
653 case '.':
654 default:
655 return 0;
659 /* Return nonzero when FNDECL represents a call to setjmp. */
662 setjmp_call_p (const_tree fndecl)
664 if (DECL_IS_RETURNS_TWICE (fndecl))
665 return ECF_RETURNS_TWICE;
666 return special_function_p (fndecl, 0) & ECF_RETURNS_TWICE;
670 /* Return true if STMT may be an alloca call. */
672 bool
673 gimple_maybe_alloca_call_p (const gimple *stmt)
675 tree fndecl;
677 if (!is_gimple_call (stmt))
678 return false;
680 fndecl = gimple_call_fndecl (stmt);
681 if (fndecl && (special_function_p (fndecl, 0) & ECF_MAY_BE_ALLOCA))
682 return true;
684 return false;
687 /* Return true if STMT is a builtin alloca call. */
689 bool
690 gimple_alloca_call_p (const gimple *stmt)
692 tree fndecl;
694 if (!is_gimple_call (stmt))
695 return false;
697 fndecl = gimple_call_fndecl (stmt);
698 if (fndecl && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
699 switch (DECL_FUNCTION_CODE (fndecl))
701 case BUILT_IN_ALLOCA:
702 case BUILT_IN_ALLOCA_WITH_ALIGN:
703 return true;
704 default:
705 break;
708 return false;
711 /* Return true when exp contains a builtin alloca call. */
713 bool
714 alloca_call_p (const_tree exp)
716 tree fndecl;
717 if (TREE_CODE (exp) == CALL_EXPR
718 && (fndecl = get_callee_fndecl (exp))
719 && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
720 switch (DECL_FUNCTION_CODE (fndecl))
722 case BUILT_IN_ALLOCA:
723 case BUILT_IN_ALLOCA_WITH_ALIGN:
724 return true;
725 default:
726 break;
729 return false;
732 /* Return TRUE if FNDECL is either a TM builtin or a TM cloned
733 function. Return FALSE otherwise. */
735 static bool
736 is_tm_builtin (const_tree fndecl)
738 if (fndecl == NULL)
739 return false;
741 if (decl_is_tm_clone (fndecl))
742 return true;
744 if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
746 switch (DECL_FUNCTION_CODE (fndecl))
748 case BUILT_IN_TM_COMMIT:
749 case BUILT_IN_TM_COMMIT_EH:
750 case BUILT_IN_TM_ABORT:
751 case BUILT_IN_TM_IRREVOCABLE:
752 case BUILT_IN_TM_GETTMCLONE_IRR:
753 case BUILT_IN_TM_MEMCPY:
754 case BUILT_IN_TM_MEMMOVE:
755 case BUILT_IN_TM_MEMSET:
756 CASE_BUILT_IN_TM_STORE (1):
757 CASE_BUILT_IN_TM_STORE (2):
758 CASE_BUILT_IN_TM_STORE (4):
759 CASE_BUILT_IN_TM_STORE (8):
760 CASE_BUILT_IN_TM_STORE (FLOAT):
761 CASE_BUILT_IN_TM_STORE (DOUBLE):
762 CASE_BUILT_IN_TM_STORE (LDOUBLE):
763 CASE_BUILT_IN_TM_STORE (M64):
764 CASE_BUILT_IN_TM_STORE (M128):
765 CASE_BUILT_IN_TM_STORE (M256):
766 CASE_BUILT_IN_TM_LOAD (1):
767 CASE_BUILT_IN_TM_LOAD (2):
768 CASE_BUILT_IN_TM_LOAD (4):
769 CASE_BUILT_IN_TM_LOAD (8):
770 CASE_BUILT_IN_TM_LOAD (FLOAT):
771 CASE_BUILT_IN_TM_LOAD (DOUBLE):
772 CASE_BUILT_IN_TM_LOAD (LDOUBLE):
773 CASE_BUILT_IN_TM_LOAD (M64):
774 CASE_BUILT_IN_TM_LOAD (M128):
775 CASE_BUILT_IN_TM_LOAD (M256):
776 case BUILT_IN_TM_LOG:
777 case BUILT_IN_TM_LOG_1:
778 case BUILT_IN_TM_LOG_2:
779 case BUILT_IN_TM_LOG_4:
780 case BUILT_IN_TM_LOG_8:
781 case BUILT_IN_TM_LOG_FLOAT:
782 case BUILT_IN_TM_LOG_DOUBLE:
783 case BUILT_IN_TM_LOG_LDOUBLE:
784 case BUILT_IN_TM_LOG_M64:
785 case BUILT_IN_TM_LOG_M128:
786 case BUILT_IN_TM_LOG_M256:
787 return true;
788 default:
789 break;
792 return false;
795 /* Detect flags (function attributes) from the function decl or type node. */
798 flags_from_decl_or_type (const_tree exp)
800 int flags = 0;
802 if (DECL_P (exp))
804 /* The function exp may have the `malloc' attribute. */
805 if (DECL_IS_MALLOC (exp))
806 flags |= ECF_MALLOC;
808 /* The function exp may have the `returns_twice' attribute. */
809 if (DECL_IS_RETURNS_TWICE (exp))
810 flags |= ECF_RETURNS_TWICE;
812 /* Process the pure and const attributes. */
813 if (TREE_READONLY (exp))
814 flags |= ECF_CONST;
815 if (DECL_PURE_P (exp))
816 flags |= ECF_PURE;
817 if (DECL_LOOPING_CONST_OR_PURE_P (exp))
818 flags |= ECF_LOOPING_CONST_OR_PURE;
820 if (DECL_IS_NOVOPS (exp))
821 flags |= ECF_NOVOPS;
822 if (lookup_attribute ("leaf", DECL_ATTRIBUTES (exp)))
823 flags |= ECF_LEAF;
825 if (TREE_NOTHROW (exp))
826 flags |= ECF_NOTHROW;
828 if (flag_tm)
830 if (is_tm_builtin (exp))
831 flags |= ECF_TM_BUILTIN;
832 else if ((flags & (ECF_CONST|ECF_NOVOPS)) != 0
833 || lookup_attribute ("transaction_pure",
834 TYPE_ATTRIBUTES (TREE_TYPE (exp))))
835 flags |= ECF_TM_PURE;
838 flags = special_function_p (exp, flags);
840 else if (TYPE_P (exp))
842 if (TYPE_READONLY (exp))
843 flags |= ECF_CONST;
845 if (flag_tm
846 && ((flags & ECF_CONST) != 0
847 || lookup_attribute ("transaction_pure", TYPE_ATTRIBUTES (exp))))
848 flags |= ECF_TM_PURE;
850 else
851 gcc_unreachable ();
853 if (TREE_THIS_VOLATILE (exp))
855 flags |= ECF_NORETURN;
856 if (flags & (ECF_CONST|ECF_PURE))
857 flags |= ECF_LOOPING_CONST_OR_PURE;
860 return flags;
863 /* Detect flags from a CALL_EXPR. */
866 call_expr_flags (const_tree t)
868 int flags;
869 tree decl = get_callee_fndecl (t);
871 if (decl)
872 flags = flags_from_decl_or_type (decl);
873 else if (CALL_EXPR_FN (t) == NULL_TREE)
874 flags = internal_fn_flags (CALL_EXPR_IFN (t));
875 else
877 tree type = TREE_TYPE (CALL_EXPR_FN (t));
878 if (type && TREE_CODE (type) == POINTER_TYPE)
879 flags = flags_from_decl_or_type (TREE_TYPE (type));
880 else
881 flags = 0;
882 if (CALL_EXPR_BY_DESCRIPTOR (t))
883 flags |= ECF_BY_DESCRIPTOR;
886 return flags;
889 /* Return true if TYPE should be passed by invisible reference. */
891 bool
892 pass_by_reference (CUMULATIVE_ARGS *ca, machine_mode mode,
893 tree type, bool named_arg)
895 if (type)
897 /* If this type contains non-trivial constructors, then it is
898 forbidden for the middle-end to create any new copies. */
899 if (TREE_ADDRESSABLE (type))
900 return true;
902 /* GCC post 3.4 passes *all* variable sized types by reference. */
903 if (!TYPE_SIZE (type) || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
904 return true;
906 /* If a record type should be passed the same as its first (and only)
907 member, use the type and mode of that member. */
908 if (TREE_CODE (type) == RECORD_TYPE && TYPE_TRANSPARENT_AGGR (type))
910 type = TREE_TYPE (first_field (type));
911 mode = TYPE_MODE (type);
915 return targetm.calls.pass_by_reference (pack_cumulative_args (ca), mode,
916 type, named_arg);
919 /* Return true if TYPE, which is passed by reference, should be callee
920 copied instead of caller copied. */
922 bool
923 reference_callee_copied (CUMULATIVE_ARGS *ca, machine_mode mode,
924 tree type, bool named_arg)
926 if (type && TREE_ADDRESSABLE (type))
927 return false;
928 return targetm.calls.callee_copies (pack_cumulative_args (ca), mode, type,
929 named_arg);
933 /* Precompute all register parameters as described by ARGS, storing values
934 into fields within the ARGS array.
936 NUM_ACTUALS indicates the total number elements in the ARGS array.
938 Set REG_PARM_SEEN if we encounter a register parameter. */
940 static void
941 precompute_register_parameters (int num_actuals, struct arg_data *args,
942 int *reg_parm_seen)
944 int i;
946 *reg_parm_seen = 0;
948 for (i = 0; i < num_actuals; i++)
949 if (args[i].reg != 0 && ! args[i].pass_on_stack)
951 *reg_parm_seen = 1;
953 if (args[i].value == 0)
955 push_temp_slots ();
956 args[i].value = expand_normal (args[i].tree_value);
957 preserve_temp_slots (args[i].value);
958 pop_temp_slots ();
961 /* If we are to promote the function arg to a wider mode,
962 do it now. */
964 if (args[i].mode != TYPE_MODE (TREE_TYPE (args[i].tree_value)))
965 args[i].value
966 = convert_modes (args[i].mode,
967 TYPE_MODE (TREE_TYPE (args[i].tree_value)),
968 args[i].value, args[i].unsignedp);
970 /* If the value is a non-legitimate constant, force it into a
971 pseudo now. TLS symbols sometimes need a call to resolve. */
972 if (CONSTANT_P (args[i].value)
973 && !targetm.legitimate_constant_p (args[i].mode, args[i].value))
974 args[i].value = force_reg (args[i].mode, args[i].value);
976 /* If we're going to have to load the value by parts, pull the
977 parts into pseudos. The part extraction process can involve
978 non-trivial computation. */
979 if (GET_CODE (args[i].reg) == PARALLEL)
981 tree type = TREE_TYPE (args[i].tree_value);
982 args[i].parallel_value
983 = emit_group_load_into_temps (args[i].reg, args[i].value,
984 type, int_size_in_bytes (type));
987 /* If the value is expensive, and we are inside an appropriately
988 short loop, put the value into a pseudo and then put the pseudo
989 into the hard reg.
991 For small register classes, also do this if this call uses
992 register parameters. This is to avoid reload conflicts while
993 loading the parameters registers. */
995 else if ((! (REG_P (args[i].value)
996 || (GET_CODE (args[i].value) == SUBREG
997 && REG_P (SUBREG_REG (args[i].value)))))
998 && args[i].mode != BLKmode
999 && (set_src_cost (args[i].value, args[i].mode,
1000 optimize_insn_for_speed_p ())
1001 > COSTS_N_INSNS (1))
1002 && ((*reg_parm_seen
1003 && targetm.small_register_classes_for_mode_p (args[i].mode))
1004 || optimize))
1005 args[i].value = copy_to_mode_reg (args[i].mode, args[i].value);
1009 #ifdef REG_PARM_STACK_SPACE
1011 /* The argument list is the property of the called routine and it
1012 may clobber it. If the fixed area has been used for previous
1013 parameters, we must save and restore it. */
1015 static rtx
1016 save_fixed_argument_area (int reg_parm_stack_space, rtx argblock, int *low_to_save, int *high_to_save)
1018 int low;
1019 int high;
1021 /* Compute the boundary of the area that needs to be saved, if any. */
1022 high = reg_parm_stack_space;
1023 if (ARGS_GROW_DOWNWARD)
1024 high += 1;
1026 if (high > highest_outgoing_arg_in_use)
1027 high = highest_outgoing_arg_in_use;
1029 for (low = 0; low < high; low++)
1030 if (stack_usage_map[low] != 0)
1032 int num_to_save;
1033 machine_mode save_mode;
1034 int delta;
1035 rtx addr;
1036 rtx stack_area;
1037 rtx save_area;
1039 while (stack_usage_map[--high] == 0)
1042 *low_to_save = low;
1043 *high_to_save = high;
1045 num_to_save = high - low + 1;
1046 save_mode = mode_for_size (num_to_save * BITS_PER_UNIT, MODE_INT, 1);
1048 /* If we don't have the required alignment, must do this
1049 in BLKmode. */
1050 if ((low & (MIN (GET_MODE_SIZE (save_mode),
1051 BIGGEST_ALIGNMENT / UNITS_PER_WORD) - 1)))
1052 save_mode = BLKmode;
1054 if (ARGS_GROW_DOWNWARD)
1055 delta = -high;
1056 else
1057 delta = low;
1059 addr = plus_constant (Pmode, argblock, delta);
1060 stack_area = gen_rtx_MEM (save_mode, memory_address (save_mode, addr));
1062 set_mem_align (stack_area, PARM_BOUNDARY);
1063 if (save_mode == BLKmode)
1065 save_area = assign_stack_temp (BLKmode, num_to_save);
1066 emit_block_move (validize_mem (save_area), stack_area,
1067 GEN_INT (num_to_save), BLOCK_OP_CALL_PARM);
1069 else
1071 save_area = gen_reg_rtx (save_mode);
1072 emit_move_insn (save_area, stack_area);
1075 return save_area;
1078 return NULL_RTX;
1081 static void
1082 restore_fixed_argument_area (rtx save_area, rtx argblock, int high_to_save, int low_to_save)
1084 machine_mode save_mode = GET_MODE (save_area);
1085 int delta;
1086 rtx addr, stack_area;
1088 if (ARGS_GROW_DOWNWARD)
1089 delta = -high_to_save;
1090 else
1091 delta = low_to_save;
1093 addr = plus_constant (Pmode, argblock, delta);
1094 stack_area = gen_rtx_MEM (save_mode, memory_address (save_mode, addr));
1095 set_mem_align (stack_area, PARM_BOUNDARY);
1097 if (save_mode != BLKmode)
1098 emit_move_insn (stack_area, save_area);
1099 else
1100 emit_block_move (stack_area, validize_mem (save_area),
1101 GEN_INT (high_to_save - low_to_save + 1),
1102 BLOCK_OP_CALL_PARM);
1104 #endif /* REG_PARM_STACK_SPACE */
1106 /* If any elements in ARGS refer to parameters that are to be passed in
1107 registers, but not in memory, and whose alignment does not permit a
1108 direct copy into registers. Copy the values into a group of pseudos
1109 which we will later copy into the appropriate hard registers.
1111 Pseudos for each unaligned argument will be stored into the array
1112 args[argnum].aligned_regs. The caller is responsible for deallocating
1113 the aligned_regs array if it is nonzero. */
1115 static void
1116 store_unaligned_arguments_into_pseudos (struct arg_data *args, int num_actuals)
1118 int i, j;
1120 for (i = 0; i < num_actuals; i++)
1121 if (args[i].reg != 0 && ! args[i].pass_on_stack
1122 && GET_CODE (args[i].reg) != PARALLEL
1123 && args[i].mode == BLKmode
1124 && MEM_P (args[i].value)
1125 && (MEM_ALIGN (args[i].value)
1126 < (unsigned int) MIN (BIGGEST_ALIGNMENT, BITS_PER_WORD)))
1128 int bytes = int_size_in_bytes (TREE_TYPE (args[i].tree_value));
1129 int endian_correction = 0;
1131 if (args[i].partial)
1133 gcc_assert (args[i].partial % UNITS_PER_WORD == 0);
1134 args[i].n_aligned_regs = args[i].partial / UNITS_PER_WORD;
1136 else
1138 args[i].n_aligned_regs
1139 = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
1142 args[i].aligned_regs = XNEWVEC (rtx, args[i].n_aligned_regs);
1144 /* Structures smaller than a word are normally aligned to the
1145 least significant byte. On a BYTES_BIG_ENDIAN machine,
1146 this means we must skip the empty high order bytes when
1147 calculating the bit offset. */
1148 if (bytes < UNITS_PER_WORD
1149 #ifdef BLOCK_REG_PADDING
1150 && (BLOCK_REG_PADDING (args[i].mode,
1151 TREE_TYPE (args[i].tree_value), 1)
1152 == downward)
1153 #else
1154 && BYTES_BIG_ENDIAN
1155 #endif
1157 endian_correction = BITS_PER_WORD - bytes * BITS_PER_UNIT;
1159 for (j = 0; j < args[i].n_aligned_regs; j++)
1161 rtx reg = gen_reg_rtx (word_mode);
1162 rtx word = operand_subword_force (args[i].value, j, BLKmode);
1163 int bitsize = MIN (bytes * BITS_PER_UNIT, BITS_PER_WORD);
1165 args[i].aligned_regs[j] = reg;
1166 word = extract_bit_field (word, bitsize, 0, 1, NULL_RTX,
1167 word_mode, word_mode, false);
1169 /* There is no need to restrict this code to loading items
1170 in TYPE_ALIGN sized hunks. The bitfield instructions can
1171 load up entire word sized registers efficiently.
1173 ??? This may not be needed anymore.
1174 We use to emit a clobber here but that doesn't let later
1175 passes optimize the instructions we emit. By storing 0 into
1176 the register later passes know the first AND to zero out the
1177 bitfield being set in the register is unnecessary. The store
1178 of 0 will be deleted as will at least the first AND. */
1180 emit_move_insn (reg, const0_rtx);
1182 bytes -= bitsize / BITS_PER_UNIT;
1183 store_bit_field (reg, bitsize, endian_correction, 0, 0,
1184 word_mode, word, false);
1189 /* The limit set by -Walloc-larger-than=. */
1190 static GTY(()) tree alloc_object_size_limit;
1192 /* Initialize ALLOC_OBJECT_SIZE_LIMIT based on the -Walloc-size-larger-than=
1193 setting if the option is specified, or to the maximum object size if it
1194 is not. Return the initialized value. */
1196 static tree
1197 alloc_max_size (void)
1199 if (!alloc_object_size_limit)
1201 alloc_object_size_limit = TYPE_MAX_VALUE (ssizetype);
1203 if (warn_alloc_size_limit)
1205 char *end = NULL;
1206 errno = 0;
1207 unsigned HOST_WIDE_INT unit = 1;
1208 unsigned HOST_WIDE_INT limit
1209 = strtoull (warn_alloc_size_limit, &end, 10);
1211 if (!errno)
1213 if (end && *end)
1215 /* Numeric option arguments are at most INT_MAX. Make it
1216 possible to specify a larger value by accepting common
1217 suffixes. */
1218 if (!strcmp (end, "kB"))
1219 unit = 1000;
1220 else if (!strcasecmp (end, "KiB") || strcmp (end, "KB"))
1221 unit = 1024;
1222 else if (!strcmp (end, "MB"))
1223 unit = 1000LU * 1000;
1224 else if (!strcasecmp (end, "MiB"))
1225 unit = 1024LU * 1024;
1226 else if (!strcasecmp (end, "GB"))
1227 unit = 1000LU * 1000 * 1000;
1228 else if (!strcasecmp (end, "GiB"))
1229 unit = 1024LU * 1024 * 1024;
1230 else if (!strcasecmp (end, "TB"))
1231 unit = 1000LU * 1000 * 1000 * 1000;
1232 else if (!strcasecmp (end, "TiB"))
1233 unit = 1024LU * 1024 * 1024 * 1024;
1234 else if (!strcasecmp (end, "PB"))
1235 unit = 1000LU * 1000 * 1000 * 1000 * 1000;
1236 else if (!strcasecmp (end, "PiB"))
1237 unit = 1024LU * 1024 * 1024 * 1024 * 1024;
1238 else if (!strcasecmp (end, "EB"))
1239 unit = 1000LU * 1000 * 1000 * 1000 * 1000 * 1000;
1240 else if (!strcasecmp (end, "EiB"))
1241 unit = 1024LU * 1024 * 1024 * 1024 * 1024 * 1024;
1242 else
1243 unit = 0;
1246 if (unit)
1247 alloc_object_size_limit
1248 = build_int_cst (ssizetype, limit * unit);
1252 return alloc_object_size_limit;
1255 /* Return true when EXP's range can be determined and set RANGE[] to it
1256 after adjusting it if necessary to make EXP a valid size argument to
1257 an allocation function declared with attribute alloc_size (whose
1258 argument may be signed), or to a string manipulation function like
1259 memset. */
1261 bool
1262 get_size_range (tree exp, tree range[2])
1264 if (tree_fits_uhwi_p (exp))
1266 /* EXP is a constant. */
1267 range[0] = range[1] = exp;
1268 return true;
1271 wide_int min, max;
1272 enum value_range_type range_type
1273 = (TREE_CODE (exp) == SSA_NAME
1274 ? get_range_info (exp, &min, &max) : VR_VARYING);
1276 if (range_type == VR_VARYING)
1278 /* No range information available. */
1279 range[0] = NULL_TREE;
1280 range[1] = NULL_TREE;
1281 return false;
1284 tree exptype = TREE_TYPE (exp);
1285 unsigned expprec = TYPE_PRECISION (exptype);
1286 wide_int wzero = wi::zero (expprec);
1287 wide_int wmaxval = wide_int (TYPE_MAX_VALUE (exptype));
1289 bool signed_p = !TYPE_UNSIGNED (exptype);
1291 if (range_type == VR_ANTI_RANGE)
1293 if (signed_p)
1295 if (wi::les_p (max, wzero))
1297 /* EXP is not in a strictly negative range. That means
1298 it must be in some (not necessarily strictly) positive
1299 range which includes zero. Since in signed to unsigned
1300 conversions negative values end up converted to large
1301 positive values, and otherwise they are not valid sizes,
1302 the resulting range is in both cases [0, TYPE_MAX]. */
1303 min = wzero;
1304 max = wmaxval;
1306 else if (wi::les_p (min - 1, wzero))
1308 /* EXP is not in a negative-positive range. That means EXP
1309 is either negative, or greater than max. Since negative
1310 sizes are invalid make the range [MAX + 1, TYPE_MAX]. */
1311 min = max + 1;
1312 max = wmaxval;
1314 else
1316 max = min - 1;
1317 min = wzero;
1320 else if (wi::eq_p (wzero, min - 1))
1322 /* EXP is unsigned and not in the range [1, MAX]. That means
1323 it's either zero or greater than MAX. Even though 0 would
1324 normally be detected by -Walloc-zero set the range to
1325 [MAX, TYPE_MAX] so that when MAX is greater than the limit
1326 the whole range is diagnosed. */
1327 min = max + 1;
1328 max = wmaxval;
1330 else
1332 max = min - 1;
1333 min = wzero;
1337 range[0] = wide_int_to_tree (exptype, min);
1338 range[1] = wide_int_to_tree (exptype, max);
1340 return true;
1343 /* Diagnose a call EXP to function FN decorated with attribute alloc_size
1344 whose argument numbers given by IDX with values given by ARGS exceed
1345 the maximum object size or cause an unsigned oveflow (wrapping) when
1346 multiplied. When ARGS[0] is null the function does nothing. ARGS[1]
1347 may be null for functions like malloc, and non-null for those like
1348 calloc that are decorated with a two-argument attribute alloc_size. */
1350 void
1351 maybe_warn_alloc_args_overflow (tree fn, tree exp, tree args[2], int idx[2])
1353 /* The range each of the (up to) two arguments is known to be in. */
1354 tree argrange[2][2] = { { NULL_TREE, NULL_TREE }, { NULL_TREE, NULL_TREE } };
1356 /* Maximum object size set by -Walloc-size-larger-than= or SIZE_MAX / 2. */
1357 tree maxobjsize = alloc_max_size ();
1359 location_t loc = EXPR_LOCATION (exp);
1361 bool warned = false;
1363 /* Validate each argument individually. */
1364 for (unsigned i = 0; i != 2 && args[i]; ++i)
1366 if (TREE_CODE (args[i]) == INTEGER_CST)
1368 argrange[i][0] = args[i];
1369 argrange[i][1] = args[i];
1371 if (tree_int_cst_lt (args[i], integer_zero_node))
1373 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1374 "%Kargument %i value %qE is negative",
1375 exp, idx[i] + 1, args[i]);
1377 else if (integer_zerop (args[i]))
1379 /* Avoid issuing -Walloc-zero for allocation functions other
1380 than __builtin_alloca that are declared with attribute
1381 returns_nonnull because there's no portability risk. This
1382 avoids warning for such calls to libiberty's xmalloc and
1383 friends.
1384 Also avoid issuing the warning for calls to function named
1385 "alloca". */
1386 if ((DECL_FUNCTION_CODE (fn) == BUILT_IN_ALLOCA
1387 && IDENTIFIER_LENGTH (DECL_NAME (fn)) != 6)
1388 || (DECL_FUNCTION_CODE (fn) != BUILT_IN_ALLOCA
1389 && !lookup_attribute ("returns_nonnull",
1390 TYPE_ATTRIBUTES (TREE_TYPE (fn)))))
1391 warned = warning_at (loc, OPT_Walloc_zero,
1392 "%Kargument %i value is zero",
1393 exp, idx[i] + 1);
1395 else if (tree_int_cst_lt (maxobjsize, args[i]))
1397 /* G++ emits calls to ::operator new[](SIZE_MAX) in C++98
1398 mode and with -fno-exceptions as a way to indicate array
1399 size overflow. There's no good way to detect C++98 here
1400 so avoid diagnosing these calls for all C++ modes. */
1401 if (i == 0
1402 && !args[1]
1403 && lang_GNU_CXX ()
1404 && DECL_IS_OPERATOR_NEW (fn)
1405 && integer_all_onesp (args[i]))
1406 continue;
1408 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1409 "%Kargument %i value %qE exceeds "
1410 "maximum object size %E",
1411 exp, idx[i] + 1, args[i], maxobjsize);
1414 else if (TREE_CODE (args[i]) == SSA_NAME
1415 && get_size_range (args[i], argrange[i]))
1417 /* Verify that the argument's range is not negative (including
1418 upper bound of zero). */
1419 if (tree_int_cst_lt (argrange[i][0], integer_zero_node)
1420 && tree_int_cst_le (argrange[i][1], integer_zero_node))
1422 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1423 "%Kargument %i range [%E, %E] is negative",
1424 exp, idx[i] + 1,
1425 argrange[i][0], argrange[i][1]);
1427 else if (tree_int_cst_lt (maxobjsize, argrange[i][0]))
1429 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1430 "%Kargument %i range [%E, %E] exceeds "
1431 "maximum object size %E",
1432 exp, idx[i] + 1,
1433 argrange[i][0], argrange[i][1],
1434 maxobjsize);
1439 if (!argrange[0])
1440 return;
1442 /* For a two-argument alloc_size, validate the product of the two
1443 arguments if both of their values or ranges are known. */
1444 if (!warned && tree_fits_uhwi_p (argrange[0][0])
1445 && argrange[1][0] && tree_fits_uhwi_p (argrange[1][0])
1446 && !integer_onep (argrange[0][0])
1447 && !integer_onep (argrange[1][0]))
1449 /* Check for overflow in the product of a function decorated with
1450 attribute alloc_size (X, Y). */
1451 unsigned szprec = TYPE_PRECISION (size_type_node);
1452 wide_int x = wi::to_wide (argrange[0][0], szprec);
1453 wide_int y = wi::to_wide (argrange[1][0], szprec);
1455 bool vflow;
1456 wide_int prod = wi::umul (x, y, &vflow);
1458 if (vflow)
1459 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1460 "%Kproduct %<%E * %E%> of arguments %i and %i "
1461 "exceeds %<SIZE_MAX%>",
1462 exp, argrange[0][0], argrange[1][0],
1463 idx[0] + 1, idx[1] + 1);
1464 else if (wi::ltu_p (wi::to_wide (maxobjsize, szprec), prod))
1465 warned = warning_at (loc, OPT_Walloc_size_larger_than_,
1466 "%Kproduct %<%E * %E%> of arguments %i and %i "
1467 "exceeds maximum object size %E",
1468 exp, argrange[0][0], argrange[1][0],
1469 idx[0] + 1, idx[1] + 1,
1470 maxobjsize);
1472 if (warned)
1474 /* Print the full range of each of the two arguments to make
1475 it clear when it is, in fact, in a range and not constant. */
1476 if (argrange[0][0] != argrange [0][1])
1477 inform (loc, "argument %i in the range [%E, %E]",
1478 idx[0] + 1, argrange[0][0], argrange[0][1]);
1479 if (argrange[1][0] != argrange [1][1])
1480 inform (loc, "argument %i in the range [%E, %E]",
1481 idx[1] + 1, argrange[1][0], argrange[1][1]);
1485 if (warned)
1487 location_t fnloc = DECL_SOURCE_LOCATION (fn);
1489 if (DECL_IS_BUILTIN (fn))
1490 inform (loc,
1491 "in a call to built-in allocation function %qD", fn);
1492 else
1493 inform (fnloc,
1494 "in a call to allocation function %qD declared here", fn);
1498 /* Issue an error if CALL_EXPR was flagged as requiring
1499 tall-call optimization. */
1501 static void
1502 maybe_complain_about_tail_call (tree call_expr, const char *reason)
1504 gcc_assert (TREE_CODE (call_expr) == CALL_EXPR);
1505 if (!CALL_EXPR_MUST_TAIL_CALL (call_expr))
1506 return;
1508 error_at (EXPR_LOCATION (call_expr), "cannot tail-call: %s", reason);
1511 /* Fill in ARGS_SIZE and ARGS array based on the parameters found in
1512 CALL_EXPR EXP.
1514 NUM_ACTUALS is the total number of parameters.
1516 N_NAMED_ARGS is the total number of named arguments.
1518 STRUCT_VALUE_ADDR_VALUE is the implicit argument for a struct return
1519 value, or null.
1521 FNDECL is the tree code for the target of this call (if known)
1523 ARGS_SO_FAR holds state needed by the target to know where to place
1524 the next argument.
1526 REG_PARM_STACK_SPACE is the number of bytes of stack space reserved
1527 for arguments which are passed in registers.
1529 OLD_STACK_LEVEL is a pointer to an rtx which olds the old stack level
1530 and may be modified by this routine.
1532 OLD_PENDING_ADJ, MUST_PREALLOCATE and FLAGS are pointers to integer
1533 flags which may be modified by this routine.
1535 MAY_TAILCALL is cleared if we encounter an invisible pass-by-reference
1536 that requires allocation of stack space.
1538 CALL_FROM_THUNK_P is true if this call is the jump from a thunk to
1539 the thunked-to function. */
1541 static void
1542 initialize_argument_information (int num_actuals ATTRIBUTE_UNUSED,
1543 struct arg_data *args,
1544 struct args_size *args_size,
1545 int n_named_args ATTRIBUTE_UNUSED,
1546 tree exp, tree struct_value_addr_value,
1547 tree fndecl, tree fntype,
1548 cumulative_args_t args_so_far,
1549 int reg_parm_stack_space,
1550 rtx *old_stack_level, int *old_pending_adj,
1551 int *must_preallocate, int *ecf_flags,
1552 bool *may_tailcall, bool call_from_thunk_p)
1554 CUMULATIVE_ARGS *args_so_far_pnt = get_cumulative_args (args_so_far);
1555 location_t loc = EXPR_LOCATION (exp);
1557 /* Count arg position in order args appear. */
1558 int argpos;
1560 int i;
1562 args_size->constant = 0;
1563 args_size->var = 0;
1565 bitmap_obstack_initialize (NULL);
1567 /* In this loop, we consider args in the order they are written.
1568 We fill up ARGS from the back. */
1570 i = num_actuals - 1;
1572 int j = i, ptr_arg = -1;
1573 call_expr_arg_iterator iter;
1574 tree arg;
1575 bitmap slots = NULL;
1577 if (struct_value_addr_value)
1579 args[j].tree_value = struct_value_addr_value;
1580 j--;
1582 /* If we pass structure address then we need to
1583 create bounds for it. Since created bounds is
1584 a call statement, we expand it right here to avoid
1585 fixing all other places where it may be expanded. */
1586 if (CALL_WITH_BOUNDS_P (exp))
1588 args[j].value = gen_reg_rtx (targetm.chkp_bound_mode ());
1589 args[j].tree_value
1590 = chkp_make_bounds_for_struct_addr (struct_value_addr_value);
1591 expand_expr_real (args[j].tree_value, args[j].value, VOIDmode,
1592 EXPAND_NORMAL, 0, false);
1593 args[j].pointer_arg = j + 1;
1594 j--;
1597 argpos = 0;
1598 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
1600 tree argtype = TREE_TYPE (arg);
1602 /* Remember last param with pointer and associate it
1603 with following pointer bounds. */
1604 if (CALL_WITH_BOUNDS_P (exp)
1605 && chkp_type_has_pointer (argtype))
1607 if (slots)
1608 BITMAP_FREE (slots);
1609 ptr_arg = j;
1610 if (!BOUNDED_TYPE_P (argtype))
1612 slots = BITMAP_ALLOC (NULL);
1613 chkp_find_bound_slots (argtype, slots);
1616 else if (CALL_WITH_BOUNDS_P (exp)
1617 && pass_by_reference (NULL, TYPE_MODE (argtype), argtype,
1618 argpos < n_named_args))
1620 if (slots)
1621 BITMAP_FREE (slots);
1622 ptr_arg = j;
1624 else if (POINTER_BOUNDS_TYPE_P (argtype))
1626 /* We expect bounds in instrumented calls only.
1627 Otherwise it is a sign we lost flag due to some optimization
1628 and may emit call args incorrectly. */
1629 gcc_assert (CALL_WITH_BOUNDS_P (exp));
1631 /* For structures look for the next available pointer. */
1632 if (ptr_arg != -1 && slots)
1634 unsigned bnd_no = bitmap_first_set_bit (slots);
1635 args[j].pointer_offset =
1636 bnd_no * POINTER_SIZE / BITS_PER_UNIT;
1638 bitmap_clear_bit (slots, bnd_no);
1640 /* Check we have no more pointers in the structure. */
1641 if (bitmap_empty_p (slots))
1642 BITMAP_FREE (slots);
1644 args[j].pointer_arg = ptr_arg;
1646 /* Check we covered all pointers in the previous
1647 non bounds arg. */
1648 if (!slots)
1649 ptr_arg = -1;
1651 else
1652 ptr_arg = -1;
1654 if (targetm.calls.split_complex_arg
1655 && argtype
1656 && TREE_CODE (argtype) == COMPLEX_TYPE
1657 && targetm.calls.split_complex_arg (argtype))
1659 tree subtype = TREE_TYPE (argtype);
1660 args[j].tree_value = build1 (REALPART_EXPR, subtype, arg);
1661 j--;
1662 args[j].tree_value = build1 (IMAGPART_EXPR, subtype, arg);
1664 else
1665 args[j].tree_value = arg;
1666 j--;
1667 argpos++;
1670 if (slots)
1671 BITMAP_FREE (slots);
1674 bitmap_obstack_release (NULL);
1676 /* Extract attribute alloc_size and if set, store the indices of
1677 the corresponding arguments in ALLOC_IDX, and then the actual
1678 argument(s) at those indices in ALLOC_ARGS. */
1679 int alloc_idx[2] = { -1, -1 };
1680 if (tree alloc_size
1681 = (fndecl ? lookup_attribute ("alloc_size",
1682 TYPE_ATTRIBUTES (TREE_TYPE (fndecl)))
1683 : NULL_TREE))
1685 tree args = TREE_VALUE (alloc_size);
1686 alloc_idx[0] = TREE_INT_CST_LOW (TREE_VALUE (args)) - 1;
1687 if (TREE_CHAIN (args))
1688 alloc_idx[1] = TREE_INT_CST_LOW (TREE_VALUE (TREE_CHAIN (args))) - 1;
1691 /* Array for up to the two attribute alloc_size arguments. */
1692 tree alloc_args[] = { NULL_TREE, NULL_TREE };
1694 /* I counts args in order (to be) pushed; ARGPOS counts in order written. */
1695 for (argpos = 0; argpos < num_actuals; i--, argpos++)
1697 tree type = TREE_TYPE (args[i].tree_value);
1698 int unsignedp;
1699 machine_mode mode;
1701 /* Replace erroneous argument with constant zero. */
1702 if (type == error_mark_node || !COMPLETE_TYPE_P (type))
1703 args[i].tree_value = integer_zero_node, type = integer_type_node;
1705 /* If TYPE is a transparent union or record, pass things the way
1706 we would pass the first field of the union or record. We have
1707 already verified that the modes are the same. */
1708 if ((TREE_CODE (type) == UNION_TYPE || TREE_CODE (type) == RECORD_TYPE)
1709 && TYPE_TRANSPARENT_AGGR (type))
1710 type = TREE_TYPE (first_field (type));
1712 /* Decide where to pass this arg.
1714 args[i].reg is nonzero if all or part is passed in registers.
1716 args[i].partial is nonzero if part but not all is passed in registers,
1717 and the exact value says how many bytes are passed in registers.
1719 args[i].pass_on_stack is nonzero if the argument must at least be
1720 computed on the stack. It may then be loaded back into registers
1721 if args[i].reg is nonzero.
1723 These decisions are driven by the FUNCTION_... macros and must agree
1724 with those made by function.c. */
1726 /* See if this argument should be passed by invisible reference. */
1727 if (pass_by_reference (args_so_far_pnt, TYPE_MODE (type),
1728 type, argpos < n_named_args))
1730 bool callee_copies;
1731 tree base = NULL_TREE;
1733 callee_copies
1734 = reference_callee_copied (args_so_far_pnt, TYPE_MODE (type),
1735 type, argpos < n_named_args);
1737 /* If we're compiling a thunk, pass through invisible references
1738 instead of making a copy. */
1739 if (call_from_thunk_p
1740 || (callee_copies
1741 && !TREE_ADDRESSABLE (type)
1742 && (base = get_base_address (args[i].tree_value))
1743 && TREE_CODE (base) != SSA_NAME
1744 && (!DECL_P (base) || MEM_P (DECL_RTL (base)))))
1746 /* We may have turned the parameter value into an SSA name.
1747 Go back to the original parameter so we can take the
1748 address. */
1749 if (TREE_CODE (args[i].tree_value) == SSA_NAME)
1751 gcc_assert (SSA_NAME_IS_DEFAULT_DEF (args[i].tree_value));
1752 args[i].tree_value = SSA_NAME_VAR (args[i].tree_value);
1753 gcc_assert (TREE_CODE (args[i].tree_value) == PARM_DECL);
1755 /* Argument setup code may have copied the value to register. We
1756 revert that optimization now because the tail call code must
1757 use the original location. */
1758 if (TREE_CODE (args[i].tree_value) == PARM_DECL
1759 && !MEM_P (DECL_RTL (args[i].tree_value))
1760 && DECL_INCOMING_RTL (args[i].tree_value)
1761 && MEM_P (DECL_INCOMING_RTL (args[i].tree_value)))
1762 set_decl_rtl (args[i].tree_value,
1763 DECL_INCOMING_RTL (args[i].tree_value));
1765 mark_addressable (args[i].tree_value);
1767 /* We can't use sibcalls if a callee-copied argument is
1768 stored in the current function's frame. */
1769 if (!call_from_thunk_p && DECL_P (base) && !TREE_STATIC (base))
1771 *may_tailcall = false;
1772 maybe_complain_about_tail_call (exp,
1773 "a callee-copied argument is"
1774 " stored in the current "
1775 " function's frame");
1778 args[i].tree_value = build_fold_addr_expr_loc (loc,
1779 args[i].tree_value);
1780 type = TREE_TYPE (args[i].tree_value);
1782 if (*ecf_flags & ECF_CONST)
1783 *ecf_flags &= ~(ECF_CONST | ECF_LOOPING_CONST_OR_PURE);
1785 else
1787 /* We make a copy of the object and pass the address to the
1788 function being called. */
1789 rtx copy;
1791 if (!COMPLETE_TYPE_P (type)
1792 || TREE_CODE (TYPE_SIZE_UNIT (type)) != INTEGER_CST
1793 || (flag_stack_check == GENERIC_STACK_CHECK
1794 && compare_tree_int (TYPE_SIZE_UNIT (type),
1795 STACK_CHECK_MAX_VAR_SIZE) > 0))
1797 /* This is a variable-sized object. Make space on the stack
1798 for it. */
1799 rtx size_rtx = expr_size (args[i].tree_value);
1801 if (*old_stack_level == 0)
1803 emit_stack_save (SAVE_BLOCK, old_stack_level);
1804 *old_pending_adj = pending_stack_adjust;
1805 pending_stack_adjust = 0;
1808 /* We can pass TRUE as the 4th argument because we just
1809 saved the stack pointer and will restore it right after
1810 the call. */
1811 copy = allocate_dynamic_stack_space (size_rtx,
1812 TYPE_ALIGN (type),
1813 TYPE_ALIGN (type),
1814 true);
1815 copy = gen_rtx_MEM (BLKmode, copy);
1816 set_mem_attributes (copy, type, 1);
1818 else
1819 copy = assign_temp (type, 1, 0);
1821 store_expr (args[i].tree_value, copy, 0, false, false);
1823 /* Just change the const function to pure and then let
1824 the next test clear the pure based on
1825 callee_copies. */
1826 if (*ecf_flags & ECF_CONST)
1828 *ecf_flags &= ~ECF_CONST;
1829 *ecf_flags |= ECF_PURE;
1832 if (!callee_copies && *ecf_flags & ECF_PURE)
1833 *ecf_flags &= ~(ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
1835 args[i].tree_value
1836 = build_fold_addr_expr_loc (loc, make_tree (type, copy));
1837 type = TREE_TYPE (args[i].tree_value);
1838 *may_tailcall = false;
1839 maybe_complain_about_tail_call (exp,
1840 "argument must be passed"
1841 " by copying");
1845 unsignedp = TYPE_UNSIGNED (type);
1846 mode = promote_function_mode (type, TYPE_MODE (type), &unsignedp,
1847 fndecl ? TREE_TYPE (fndecl) : fntype, 0);
1849 args[i].unsignedp = unsignedp;
1850 args[i].mode = mode;
1852 args[i].reg = targetm.calls.function_arg (args_so_far, mode, type,
1853 argpos < n_named_args);
1855 if (args[i].reg && CONST_INT_P (args[i].reg))
1857 args[i].special_slot = args[i].reg;
1858 args[i].reg = NULL;
1861 /* If this is a sibling call and the machine has register windows, the
1862 register window has to be unwinded before calling the routine, so
1863 arguments have to go into the incoming registers. */
1864 if (targetm.calls.function_incoming_arg != targetm.calls.function_arg)
1865 args[i].tail_call_reg
1866 = targetm.calls.function_incoming_arg (args_so_far, mode, type,
1867 argpos < n_named_args);
1868 else
1869 args[i].tail_call_reg = args[i].reg;
1871 if (args[i].reg)
1872 args[i].partial
1873 = targetm.calls.arg_partial_bytes (args_so_far, mode, type,
1874 argpos < n_named_args);
1876 args[i].pass_on_stack = targetm.calls.must_pass_in_stack (mode, type);
1878 /* If FUNCTION_ARG returned a (parallel [(expr_list (nil) ...) ...]),
1879 it means that we are to pass this arg in the register(s) designated
1880 by the PARALLEL, but also to pass it in the stack. */
1881 if (args[i].reg && GET_CODE (args[i].reg) == PARALLEL
1882 && XEXP (XVECEXP (args[i].reg, 0, 0), 0) == 0)
1883 args[i].pass_on_stack = 1;
1885 /* If this is an addressable type, we must preallocate the stack
1886 since we must evaluate the object into its final location.
1888 If this is to be passed in both registers and the stack, it is simpler
1889 to preallocate. */
1890 if (TREE_ADDRESSABLE (type)
1891 || (args[i].pass_on_stack && args[i].reg != 0))
1892 *must_preallocate = 1;
1894 /* No stack allocation and padding for bounds. */
1895 if (POINTER_BOUNDS_P (args[i].tree_value))
1897 /* Compute the stack-size of this argument. */
1898 else if (args[i].reg == 0 || args[i].partial != 0
1899 || reg_parm_stack_space > 0
1900 || args[i].pass_on_stack)
1901 locate_and_pad_parm (mode, type,
1902 #ifdef STACK_PARMS_IN_REG_PARM_AREA
1904 #else
1905 args[i].reg != 0,
1906 #endif
1907 reg_parm_stack_space,
1908 args[i].pass_on_stack ? 0 : args[i].partial,
1909 fndecl, args_size, &args[i].locate);
1910 #ifdef BLOCK_REG_PADDING
1911 else
1912 /* The argument is passed entirely in registers. See at which
1913 end it should be padded. */
1914 args[i].locate.where_pad =
1915 BLOCK_REG_PADDING (mode, type,
1916 int_size_in_bytes (type) <= UNITS_PER_WORD);
1917 #endif
1919 /* Update ARGS_SIZE, the total stack space for args so far. */
1921 args_size->constant += args[i].locate.size.constant;
1922 if (args[i].locate.size.var)
1923 ADD_PARM_SIZE (*args_size, args[i].locate.size.var);
1925 /* Increment ARGS_SO_FAR, which has info about which arg-registers
1926 have been used, etc. */
1928 targetm.calls.function_arg_advance (args_so_far, TYPE_MODE (type),
1929 type, argpos < n_named_args);
1931 /* Store argument values for functions decorated with attribute
1932 alloc_size. */
1933 if (argpos == alloc_idx[0])
1934 alloc_args[0] = args[i].tree_value;
1935 else if (argpos == alloc_idx[1])
1936 alloc_args[1] = args[i].tree_value;
1939 if (alloc_args[0])
1941 /* Check the arguments of functions decorated with attribute
1942 alloc_size. */
1943 maybe_warn_alloc_args_overflow (fndecl, exp, alloc_args, alloc_idx);
1947 /* Update ARGS_SIZE to contain the total size for the argument block.
1948 Return the original constant component of the argument block's size.
1950 REG_PARM_STACK_SPACE holds the number of bytes of stack space reserved
1951 for arguments passed in registers. */
1953 static int
1954 compute_argument_block_size (int reg_parm_stack_space,
1955 struct args_size *args_size,
1956 tree fndecl ATTRIBUTE_UNUSED,
1957 tree fntype ATTRIBUTE_UNUSED,
1958 int preferred_stack_boundary ATTRIBUTE_UNUSED)
1960 int unadjusted_args_size = args_size->constant;
1962 /* For accumulate outgoing args mode we don't need to align, since the frame
1963 will be already aligned. Align to STACK_BOUNDARY in order to prevent
1964 backends from generating misaligned frame sizes. */
1965 if (ACCUMULATE_OUTGOING_ARGS && preferred_stack_boundary > STACK_BOUNDARY)
1966 preferred_stack_boundary = STACK_BOUNDARY;
1968 /* Compute the actual size of the argument block required. The variable
1969 and constant sizes must be combined, the size may have to be rounded,
1970 and there may be a minimum required size. */
1972 if (args_size->var)
1974 args_size->var = ARGS_SIZE_TREE (*args_size);
1975 args_size->constant = 0;
1977 preferred_stack_boundary /= BITS_PER_UNIT;
1978 if (preferred_stack_boundary > 1)
1980 /* We don't handle this case yet. To handle it correctly we have
1981 to add the delta, round and subtract the delta.
1982 Currently no machine description requires this support. */
1983 gcc_assert (!(stack_pointer_delta & (preferred_stack_boundary - 1)));
1984 args_size->var = round_up (args_size->var, preferred_stack_boundary);
1987 if (reg_parm_stack_space > 0)
1989 args_size->var
1990 = size_binop (MAX_EXPR, args_size->var,
1991 ssize_int (reg_parm_stack_space));
1993 /* The area corresponding to register parameters is not to count in
1994 the size of the block we need. So make the adjustment. */
1995 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
1996 args_size->var
1997 = size_binop (MINUS_EXPR, args_size->var,
1998 ssize_int (reg_parm_stack_space));
2001 else
2003 preferred_stack_boundary /= BITS_PER_UNIT;
2004 if (preferred_stack_boundary < 1)
2005 preferred_stack_boundary = 1;
2006 args_size->constant = (((args_size->constant
2007 + stack_pointer_delta
2008 + preferred_stack_boundary - 1)
2009 / preferred_stack_boundary
2010 * preferred_stack_boundary)
2011 - stack_pointer_delta);
2013 args_size->constant = MAX (args_size->constant,
2014 reg_parm_stack_space);
2016 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
2017 args_size->constant -= reg_parm_stack_space;
2019 return unadjusted_args_size;
2022 /* Precompute parameters as needed for a function call.
2024 FLAGS is mask of ECF_* constants.
2026 NUM_ACTUALS is the number of arguments.
2028 ARGS is an array containing information for each argument; this
2029 routine fills in the INITIAL_VALUE and VALUE fields for each
2030 precomputed argument. */
2032 static void
2033 precompute_arguments (int num_actuals, struct arg_data *args)
2035 int i;
2037 /* If this is a libcall, then precompute all arguments so that we do not
2038 get extraneous instructions emitted as part of the libcall sequence. */
2040 /* If we preallocated the stack space, and some arguments must be passed
2041 on the stack, then we must precompute any parameter which contains a
2042 function call which will store arguments on the stack.
2043 Otherwise, evaluating the parameter may clobber previous parameters
2044 which have already been stored into the stack. (we have code to avoid
2045 such case by saving the outgoing stack arguments, but it results in
2046 worse code) */
2047 if (!ACCUMULATE_OUTGOING_ARGS)
2048 return;
2050 for (i = 0; i < num_actuals; i++)
2052 tree type;
2053 machine_mode mode;
2055 if (TREE_CODE (args[i].tree_value) != CALL_EXPR)
2056 continue;
2058 /* If this is an addressable type, we cannot pre-evaluate it. */
2059 type = TREE_TYPE (args[i].tree_value);
2060 gcc_assert (!TREE_ADDRESSABLE (type));
2062 args[i].initial_value = args[i].value
2063 = expand_normal (args[i].tree_value);
2065 mode = TYPE_MODE (type);
2066 if (mode != args[i].mode)
2068 int unsignedp = args[i].unsignedp;
2069 args[i].value
2070 = convert_modes (args[i].mode, mode,
2071 args[i].value, args[i].unsignedp);
2073 /* CSE will replace this only if it contains args[i].value
2074 pseudo, so convert it down to the declared mode using
2075 a SUBREG. */
2076 if (REG_P (args[i].value)
2077 && GET_MODE_CLASS (args[i].mode) == MODE_INT
2078 && promote_mode (type, mode, &unsignedp) != args[i].mode)
2080 args[i].initial_value
2081 = gen_lowpart_SUBREG (mode, args[i].value);
2082 SUBREG_PROMOTED_VAR_P (args[i].initial_value) = 1;
2083 SUBREG_PROMOTED_SET (args[i].initial_value, args[i].unsignedp);
2089 /* Given the current state of MUST_PREALLOCATE and information about
2090 arguments to a function call in NUM_ACTUALS, ARGS and ARGS_SIZE,
2091 compute and return the final value for MUST_PREALLOCATE. */
2093 static int
2094 finalize_must_preallocate (int must_preallocate, int num_actuals,
2095 struct arg_data *args, struct args_size *args_size)
2097 /* See if we have or want to preallocate stack space.
2099 If we would have to push a partially-in-regs parm
2100 before other stack parms, preallocate stack space instead.
2102 If the size of some parm is not a multiple of the required stack
2103 alignment, we must preallocate.
2105 If the total size of arguments that would otherwise create a copy in
2106 a temporary (such as a CALL) is more than half the total argument list
2107 size, preallocation is faster.
2109 Another reason to preallocate is if we have a machine (like the m88k)
2110 where stack alignment is required to be maintained between every
2111 pair of insns, not just when the call is made. However, we assume here
2112 that such machines either do not have push insns (and hence preallocation
2113 would occur anyway) or the problem is taken care of with
2114 PUSH_ROUNDING. */
2116 if (! must_preallocate)
2118 int partial_seen = 0;
2119 int copy_to_evaluate_size = 0;
2120 int i;
2122 for (i = 0; i < num_actuals && ! must_preallocate; i++)
2124 if (args[i].partial > 0 && ! args[i].pass_on_stack)
2125 partial_seen = 1;
2126 else if (partial_seen && args[i].reg == 0)
2127 must_preallocate = 1;
2128 /* We preallocate in case there are bounds passed
2129 in the bounds table to have precomputed address
2130 for bounds association. */
2131 else if (POINTER_BOUNDS_P (args[i].tree_value)
2132 && !args[i].reg)
2133 must_preallocate = 1;
2135 if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode
2136 && (TREE_CODE (args[i].tree_value) == CALL_EXPR
2137 || TREE_CODE (args[i].tree_value) == TARGET_EXPR
2138 || TREE_CODE (args[i].tree_value) == COND_EXPR
2139 || TREE_ADDRESSABLE (TREE_TYPE (args[i].tree_value))))
2140 copy_to_evaluate_size
2141 += int_size_in_bytes (TREE_TYPE (args[i].tree_value));
2144 if (copy_to_evaluate_size * 2 >= args_size->constant
2145 && args_size->constant > 0)
2146 must_preallocate = 1;
2148 return must_preallocate;
2151 /* If we preallocated stack space, compute the address of each argument
2152 and store it into the ARGS array.
2154 We need not ensure it is a valid memory address here; it will be
2155 validized when it is used.
2157 ARGBLOCK is an rtx for the address of the outgoing arguments. */
2159 static void
2160 compute_argument_addresses (struct arg_data *args, rtx argblock, int num_actuals)
2162 if (argblock)
2164 rtx arg_reg = argblock;
2165 int i, arg_offset = 0;
2167 if (GET_CODE (argblock) == PLUS)
2168 arg_reg = XEXP (argblock, 0), arg_offset = INTVAL (XEXP (argblock, 1));
2170 for (i = 0; i < num_actuals; i++)
2172 rtx offset = ARGS_SIZE_RTX (args[i].locate.offset);
2173 rtx slot_offset = ARGS_SIZE_RTX (args[i].locate.slot_offset);
2174 rtx addr;
2175 unsigned int align, boundary;
2176 unsigned int units_on_stack = 0;
2177 machine_mode partial_mode = VOIDmode;
2179 /* Skip this parm if it will not be passed on the stack. */
2180 if (! args[i].pass_on_stack
2181 && args[i].reg != 0
2182 && args[i].partial == 0)
2183 continue;
2185 /* Pointer Bounds are never passed on the stack. */
2186 if (POINTER_BOUNDS_P (args[i].tree_value))
2187 continue;
2189 if (CONST_INT_P (offset))
2190 addr = plus_constant (Pmode, arg_reg, INTVAL (offset));
2191 else
2192 addr = gen_rtx_PLUS (Pmode, arg_reg, offset);
2194 addr = plus_constant (Pmode, addr, arg_offset);
2196 if (args[i].partial != 0)
2198 /* Only part of the parameter is being passed on the stack.
2199 Generate a simple memory reference of the correct size. */
2200 units_on_stack = args[i].locate.size.constant;
2201 partial_mode = mode_for_size (units_on_stack * BITS_PER_UNIT,
2202 MODE_INT, 1);
2203 args[i].stack = gen_rtx_MEM (partial_mode, addr);
2204 set_mem_size (args[i].stack, units_on_stack);
2206 else
2208 args[i].stack = gen_rtx_MEM (args[i].mode, addr);
2209 set_mem_attributes (args[i].stack,
2210 TREE_TYPE (args[i].tree_value), 1);
2212 align = BITS_PER_UNIT;
2213 boundary = args[i].locate.boundary;
2214 if (args[i].locate.where_pad != downward)
2215 align = boundary;
2216 else if (CONST_INT_P (offset))
2218 align = INTVAL (offset) * BITS_PER_UNIT | boundary;
2219 align = least_bit_hwi (align);
2221 set_mem_align (args[i].stack, align);
2223 if (CONST_INT_P (slot_offset))
2224 addr = plus_constant (Pmode, arg_reg, INTVAL (slot_offset));
2225 else
2226 addr = gen_rtx_PLUS (Pmode, arg_reg, slot_offset);
2228 addr = plus_constant (Pmode, addr, arg_offset);
2230 if (args[i].partial != 0)
2232 /* Only part of the parameter is being passed on the stack.
2233 Generate a simple memory reference of the correct size.
2235 args[i].stack_slot = gen_rtx_MEM (partial_mode, addr);
2236 set_mem_size (args[i].stack_slot, units_on_stack);
2238 else
2240 args[i].stack_slot = gen_rtx_MEM (args[i].mode, addr);
2241 set_mem_attributes (args[i].stack_slot,
2242 TREE_TYPE (args[i].tree_value), 1);
2244 set_mem_align (args[i].stack_slot, args[i].locate.boundary);
2246 /* Function incoming arguments may overlap with sibling call
2247 outgoing arguments and we cannot allow reordering of reads
2248 from function arguments with stores to outgoing arguments
2249 of sibling calls. */
2250 set_mem_alias_set (args[i].stack, 0);
2251 set_mem_alias_set (args[i].stack_slot, 0);
2256 /* Given a FNDECL and EXP, return an rtx suitable for use as a target address
2257 in a call instruction.
2259 FNDECL is the tree node for the target function. For an indirect call
2260 FNDECL will be NULL_TREE.
2262 ADDR is the operand 0 of CALL_EXPR for this call. */
2264 static rtx
2265 rtx_for_function_call (tree fndecl, tree addr)
2267 rtx funexp;
2269 /* Get the function to call, in the form of RTL. */
2270 if (fndecl)
2272 if (!TREE_USED (fndecl) && fndecl != current_function_decl)
2273 TREE_USED (fndecl) = 1;
2275 /* Get a SYMBOL_REF rtx for the function address. */
2276 funexp = XEXP (DECL_RTL (fndecl), 0);
2278 else
2279 /* Generate an rtx (probably a pseudo-register) for the address. */
2281 push_temp_slots ();
2282 funexp = expand_normal (addr);
2283 pop_temp_slots (); /* FUNEXP can't be BLKmode. */
2285 return funexp;
2288 /* Internal state for internal_arg_pointer_based_exp and its helpers. */
2289 static struct
2291 /* Last insn that has been scanned by internal_arg_pointer_based_exp_scan,
2292 or NULL_RTX if none has been scanned yet. */
2293 rtx_insn *scan_start;
2294 /* Vector indexed by REGNO - FIRST_PSEUDO_REGISTER, recording if a pseudo is
2295 based on crtl->args.internal_arg_pointer. The element is NULL_RTX if the
2296 pseudo isn't based on it, a CONST_INT offset if the pseudo is based on it
2297 with fixed offset, or PC if this is with variable or unknown offset. */
2298 vec<rtx> cache;
2299 } internal_arg_pointer_exp_state;
2301 static rtx internal_arg_pointer_based_exp (const_rtx, bool);
2303 /* Helper function for internal_arg_pointer_based_exp. Scan insns in
2304 the tail call sequence, starting with first insn that hasn't been
2305 scanned yet, and note for each pseudo on the LHS whether it is based
2306 on crtl->args.internal_arg_pointer or not, and what offset from that
2307 that pointer it has. */
2309 static void
2310 internal_arg_pointer_based_exp_scan (void)
2312 rtx_insn *insn, *scan_start = internal_arg_pointer_exp_state.scan_start;
2314 if (scan_start == NULL_RTX)
2315 insn = get_insns ();
2316 else
2317 insn = NEXT_INSN (scan_start);
2319 while (insn)
2321 rtx set = single_set (insn);
2322 if (set && REG_P (SET_DEST (set)) && !HARD_REGISTER_P (SET_DEST (set)))
2324 rtx val = NULL_RTX;
2325 unsigned int idx = REGNO (SET_DEST (set)) - FIRST_PSEUDO_REGISTER;
2326 /* Punt on pseudos set multiple times. */
2327 if (idx < internal_arg_pointer_exp_state.cache.length ()
2328 && (internal_arg_pointer_exp_state.cache[idx]
2329 != NULL_RTX))
2330 val = pc_rtx;
2331 else
2332 val = internal_arg_pointer_based_exp (SET_SRC (set), false);
2333 if (val != NULL_RTX)
2335 if (idx >= internal_arg_pointer_exp_state.cache.length ())
2336 internal_arg_pointer_exp_state.cache
2337 .safe_grow_cleared (idx + 1);
2338 internal_arg_pointer_exp_state.cache[idx] = val;
2341 if (NEXT_INSN (insn) == NULL_RTX)
2342 scan_start = insn;
2343 insn = NEXT_INSN (insn);
2346 internal_arg_pointer_exp_state.scan_start = scan_start;
2349 /* Compute whether RTL is based on crtl->args.internal_arg_pointer. Return
2350 NULL_RTX if RTL isn't based on it, a CONST_INT offset if RTL is based on
2351 it with fixed offset, or PC if this is with variable or unknown offset.
2352 TOPLEVEL is true if the function is invoked at the topmost level. */
2354 static rtx
2355 internal_arg_pointer_based_exp (const_rtx rtl, bool toplevel)
2357 if (CONSTANT_P (rtl))
2358 return NULL_RTX;
2360 if (rtl == crtl->args.internal_arg_pointer)
2361 return const0_rtx;
2363 if (REG_P (rtl) && HARD_REGISTER_P (rtl))
2364 return NULL_RTX;
2366 if (GET_CODE (rtl) == PLUS && CONST_INT_P (XEXP (rtl, 1)))
2368 rtx val = internal_arg_pointer_based_exp (XEXP (rtl, 0), toplevel);
2369 if (val == NULL_RTX || val == pc_rtx)
2370 return val;
2371 return plus_constant (Pmode, val, INTVAL (XEXP (rtl, 1)));
2374 /* When called at the topmost level, scan pseudo assignments in between the
2375 last scanned instruction in the tail call sequence and the latest insn
2376 in that sequence. */
2377 if (toplevel)
2378 internal_arg_pointer_based_exp_scan ();
2380 if (REG_P (rtl))
2382 unsigned int idx = REGNO (rtl) - FIRST_PSEUDO_REGISTER;
2383 if (idx < internal_arg_pointer_exp_state.cache.length ())
2384 return internal_arg_pointer_exp_state.cache[idx];
2386 return NULL_RTX;
2389 subrtx_iterator::array_type array;
2390 FOR_EACH_SUBRTX (iter, array, rtl, NONCONST)
2392 const_rtx x = *iter;
2393 if (REG_P (x) && internal_arg_pointer_based_exp (x, false) != NULL_RTX)
2394 return pc_rtx;
2395 if (MEM_P (x))
2396 iter.skip_subrtxes ();
2399 return NULL_RTX;
2402 /* Return true if and only if SIZE storage units (usually bytes)
2403 starting from address ADDR overlap with already clobbered argument
2404 area. This function is used to determine if we should give up a
2405 sibcall. */
2407 static bool
2408 mem_overlaps_already_clobbered_arg_p (rtx addr, unsigned HOST_WIDE_INT size)
2410 HOST_WIDE_INT i;
2411 rtx val;
2413 if (bitmap_empty_p (stored_args_map))
2414 return false;
2415 val = internal_arg_pointer_based_exp (addr, true);
2416 if (val == NULL_RTX)
2417 return false;
2418 else if (val == pc_rtx)
2419 return true;
2420 else
2421 i = INTVAL (val);
2423 if (STACK_GROWS_DOWNWARD)
2424 i -= crtl->args.pretend_args_size;
2425 else
2426 i += crtl->args.pretend_args_size;
2429 if (ARGS_GROW_DOWNWARD)
2430 i = -i - size;
2432 if (size > 0)
2434 unsigned HOST_WIDE_INT k;
2436 for (k = 0; k < size; k++)
2437 if (i + k < SBITMAP_SIZE (stored_args_map)
2438 && bitmap_bit_p (stored_args_map, i + k))
2439 return true;
2442 return false;
2445 /* Do the register loads required for any wholly-register parms or any
2446 parms which are passed both on the stack and in a register. Their
2447 expressions were already evaluated.
2449 Mark all register-parms as living through the call, putting these USE
2450 insns in the CALL_INSN_FUNCTION_USAGE field.
2452 When IS_SIBCALL, perform the check_sibcall_argument_overlap
2453 checking, setting *SIBCALL_FAILURE if appropriate. */
2455 static void
2456 load_register_parameters (struct arg_data *args, int num_actuals,
2457 rtx *call_fusage, int flags, int is_sibcall,
2458 int *sibcall_failure)
2460 int i, j;
2462 for (i = 0; i < num_actuals; i++)
2464 rtx reg = ((flags & ECF_SIBCALL)
2465 ? args[i].tail_call_reg : args[i].reg);
2466 if (reg)
2468 int partial = args[i].partial;
2469 int nregs;
2470 int size = 0;
2471 rtx_insn *before_arg = get_last_insn ();
2472 /* Set non-negative if we must move a word at a time, even if
2473 just one word (e.g, partial == 4 && mode == DFmode). Set
2474 to -1 if we just use a normal move insn. This value can be
2475 zero if the argument is a zero size structure. */
2476 nregs = -1;
2477 if (GET_CODE (reg) == PARALLEL)
2479 else if (partial)
2481 gcc_assert (partial % UNITS_PER_WORD == 0);
2482 nregs = partial / UNITS_PER_WORD;
2484 else if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode)
2486 size = int_size_in_bytes (TREE_TYPE (args[i].tree_value));
2487 nregs = (size + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD;
2489 else
2490 size = GET_MODE_SIZE (args[i].mode);
2492 /* Handle calls that pass values in multiple non-contiguous
2493 locations. The Irix 6 ABI has examples of this. */
2495 if (GET_CODE (reg) == PARALLEL)
2496 emit_group_move (reg, args[i].parallel_value);
2498 /* If simple case, just do move. If normal partial, store_one_arg
2499 has already loaded the register for us. In all other cases,
2500 load the register(s) from memory. */
2502 else if (nregs == -1)
2504 emit_move_insn (reg, args[i].value);
2505 #ifdef BLOCK_REG_PADDING
2506 /* Handle case where we have a value that needs shifting
2507 up to the msb. eg. a QImode value and we're padding
2508 upward on a BYTES_BIG_ENDIAN machine. */
2509 if (size < UNITS_PER_WORD
2510 && (args[i].locate.where_pad
2511 == (BYTES_BIG_ENDIAN ? upward : downward)))
2513 rtx x;
2514 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
2516 /* Assigning REG here rather than a temp makes CALL_FUSAGE
2517 report the whole reg as used. Strictly speaking, the
2518 call only uses SIZE bytes at the msb end, but it doesn't
2519 seem worth generating rtl to say that. */
2520 reg = gen_rtx_REG (word_mode, REGNO (reg));
2521 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1);
2522 if (x != reg)
2523 emit_move_insn (reg, x);
2525 #endif
2528 /* If we have pre-computed the values to put in the registers in
2529 the case of non-aligned structures, copy them in now. */
2531 else if (args[i].n_aligned_regs != 0)
2532 for (j = 0; j < args[i].n_aligned_regs; j++)
2533 emit_move_insn (gen_rtx_REG (word_mode, REGNO (reg) + j),
2534 args[i].aligned_regs[j]);
2536 else if (partial == 0 || args[i].pass_on_stack)
2538 rtx mem = validize_mem (copy_rtx (args[i].value));
2540 /* Check for overlap with already clobbered argument area,
2541 providing that this has non-zero size. */
2542 if (is_sibcall
2543 && size != 0
2544 && (mem_overlaps_already_clobbered_arg_p
2545 (XEXP (args[i].value, 0), size)))
2546 *sibcall_failure = 1;
2548 if (size % UNITS_PER_WORD == 0
2549 || MEM_ALIGN (mem) % BITS_PER_WORD == 0)
2550 move_block_to_reg (REGNO (reg), mem, nregs, args[i].mode);
2551 else
2553 if (nregs > 1)
2554 move_block_to_reg (REGNO (reg), mem, nregs - 1,
2555 args[i].mode);
2556 rtx dest = gen_rtx_REG (word_mode, REGNO (reg) + nregs - 1);
2557 unsigned int bitoff = (nregs - 1) * BITS_PER_WORD;
2558 unsigned int bitsize = size * BITS_PER_UNIT - bitoff;
2559 rtx x = extract_bit_field (mem, bitsize, bitoff, 1, dest,
2560 word_mode, word_mode, false);
2561 if (BYTES_BIG_ENDIAN)
2562 x = expand_shift (LSHIFT_EXPR, word_mode, x,
2563 BITS_PER_WORD - bitsize, dest, 1);
2564 if (x != dest)
2565 emit_move_insn (dest, x);
2568 /* Handle a BLKmode that needs shifting. */
2569 if (nregs == 1 && size < UNITS_PER_WORD
2570 #ifdef BLOCK_REG_PADDING
2571 && args[i].locate.where_pad == downward
2572 #else
2573 && BYTES_BIG_ENDIAN
2574 #endif
2577 rtx dest = gen_rtx_REG (word_mode, REGNO (reg));
2578 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
2579 enum tree_code dir = (BYTES_BIG_ENDIAN
2580 ? RSHIFT_EXPR : LSHIFT_EXPR);
2581 rtx x;
2583 x = expand_shift (dir, word_mode, dest, shift, dest, 1);
2584 if (x != dest)
2585 emit_move_insn (dest, x);
2589 /* When a parameter is a block, and perhaps in other cases, it is
2590 possible that it did a load from an argument slot that was
2591 already clobbered. */
2592 if (is_sibcall
2593 && check_sibcall_argument_overlap (before_arg, &args[i], 0))
2594 *sibcall_failure = 1;
2596 /* Handle calls that pass values in multiple non-contiguous
2597 locations. The Irix 6 ABI has examples of this. */
2598 if (GET_CODE (reg) == PARALLEL)
2599 use_group_regs (call_fusage, reg);
2600 else if (nregs == -1)
2601 use_reg_mode (call_fusage, reg,
2602 TYPE_MODE (TREE_TYPE (args[i].tree_value)));
2603 else if (nregs > 0)
2604 use_regs (call_fusage, REGNO (reg), nregs);
2609 /* We need to pop PENDING_STACK_ADJUST bytes. But, if the arguments
2610 wouldn't fill up an even multiple of PREFERRED_UNIT_STACK_BOUNDARY
2611 bytes, then we would need to push some additional bytes to pad the
2612 arguments. So, we compute an adjust to the stack pointer for an
2613 amount that will leave the stack under-aligned by UNADJUSTED_ARGS_SIZE
2614 bytes. Then, when the arguments are pushed the stack will be perfectly
2615 aligned. ARGS_SIZE->CONSTANT is set to the number of bytes that should
2616 be popped after the call. Returns the adjustment. */
2618 static int
2619 combine_pending_stack_adjustment_and_call (int unadjusted_args_size,
2620 struct args_size *args_size,
2621 unsigned int preferred_unit_stack_boundary)
2623 /* The number of bytes to pop so that the stack will be
2624 under-aligned by UNADJUSTED_ARGS_SIZE bytes. */
2625 HOST_WIDE_INT adjustment;
2626 /* The alignment of the stack after the arguments are pushed, if we
2627 just pushed the arguments without adjust the stack here. */
2628 unsigned HOST_WIDE_INT unadjusted_alignment;
2630 unadjusted_alignment
2631 = ((stack_pointer_delta + unadjusted_args_size)
2632 % preferred_unit_stack_boundary);
2634 /* We want to get rid of as many of the PENDING_STACK_ADJUST bytes
2635 as possible -- leaving just enough left to cancel out the
2636 UNADJUSTED_ALIGNMENT. In other words, we want to ensure that the
2637 PENDING_STACK_ADJUST is non-negative, and congruent to
2638 -UNADJUSTED_ALIGNMENT modulo the PREFERRED_UNIT_STACK_BOUNDARY. */
2640 /* Begin by trying to pop all the bytes. */
2641 unadjusted_alignment
2642 = (unadjusted_alignment
2643 - (pending_stack_adjust % preferred_unit_stack_boundary));
2644 adjustment = pending_stack_adjust;
2645 /* Push enough additional bytes that the stack will be aligned
2646 after the arguments are pushed. */
2647 if (preferred_unit_stack_boundary > 1 && unadjusted_alignment)
2648 adjustment -= preferred_unit_stack_boundary - unadjusted_alignment;
2650 /* Now, sets ARGS_SIZE->CONSTANT so that we pop the right number of
2651 bytes after the call. The right number is the entire
2652 PENDING_STACK_ADJUST less our ADJUSTMENT plus the amount required
2653 by the arguments in the first place. */
2654 args_size->constant
2655 = pending_stack_adjust - adjustment + unadjusted_args_size;
2657 return adjustment;
2660 /* Scan X expression if it does not dereference any argument slots
2661 we already clobbered by tail call arguments (as noted in stored_args_map
2662 bitmap).
2663 Return nonzero if X expression dereferences such argument slots,
2664 zero otherwise. */
2666 static int
2667 check_sibcall_argument_overlap_1 (rtx x)
2669 RTX_CODE code;
2670 int i, j;
2671 const char *fmt;
2673 if (x == NULL_RTX)
2674 return 0;
2676 code = GET_CODE (x);
2678 /* We need not check the operands of the CALL expression itself. */
2679 if (code == CALL)
2680 return 0;
2682 if (code == MEM)
2683 return mem_overlaps_already_clobbered_arg_p (XEXP (x, 0),
2684 GET_MODE_SIZE (GET_MODE (x)));
2686 /* Scan all subexpressions. */
2687 fmt = GET_RTX_FORMAT (code);
2688 for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++)
2690 if (*fmt == 'e')
2692 if (check_sibcall_argument_overlap_1 (XEXP (x, i)))
2693 return 1;
2695 else if (*fmt == 'E')
2697 for (j = 0; j < XVECLEN (x, i); j++)
2698 if (check_sibcall_argument_overlap_1 (XVECEXP (x, i, j)))
2699 return 1;
2702 return 0;
2705 /* Scan sequence after INSN if it does not dereference any argument slots
2706 we already clobbered by tail call arguments (as noted in stored_args_map
2707 bitmap). If MARK_STORED_ARGS_MAP, add stack slots for ARG to
2708 stored_args_map bitmap afterwards (when ARG is a register MARK_STORED_ARGS_MAP
2709 should be 0). Return nonzero if sequence after INSN dereferences such argument
2710 slots, zero otherwise. */
2712 static int
2713 check_sibcall_argument_overlap (rtx_insn *insn, struct arg_data *arg,
2714 int mark_stored_args_map)
2716 int low, high;
2718 if (insn == NULL_RTX)
2719 insn = get_insns ();
2720 else
2721 insn = NEXT_INSN (insn);
2723 for (; insn; insn = NEXT_INSN (insn))
2724 if (INSN_P (insn)
2725 && check_sibcall_argument_overlap_1 (PATTERN (insn)))
2726 break;
2728 if (mark_stored_args_map)
2730 if (ARGS_GROW_DOWNWARD)
2731 low = -arg->locate.slot_offset.constant - arg->locate.size.constant;
2732 else
2733 low = arg->locate.slot_offset.constant;
2735 for (high = low + arg->locate.size.constant; low < high; low++)
2736 bitmap_set_bit (stored_args_map, low);
2738 return insn != NULL_RTX;
2741 /* Given that a function returns a value of mode MODE at the most
2742 significant end of hard register VALUE, shift VALUE left or right
2743 as specified by LEFT_P. Return true if some action was needed. */
2745 bool
2746 shift_return_value (machine_mode mode, bool left_p, rtx value)
2748 HOST_WIDE_INT shift;
2750 gcc_assert (REG_P (value) && HARD_REGISTER_P (value));
2751 shift = GET_MODE_BITSIZE (GET_MODE (value)) - GET_MODE_BITSIZE (mode);
2752 if (shift == 0)
2753 return false;
2755 /* Use ashr rather than lshr for right shifts. This is for the benefit
2756 of the MIPS port, which requires SImode values to be sign-extended
2757 when stored in 64-bit registers. */
2758 if (!force_expand_binop (GET_MODE (value), left_p ? ashl_optab : ashr_optab,
2759 value, GEN_INT (shift), value, 1, OPTAB_WIDEN))
2760 gcc_unreachable ();
2761 return true;
2764 /* If X is a likely-spilled register value, copy it to a pseudo
2765 register and return that register. Return X otherwise. */
2767 static rtx
2768 avoid_likely_spilled_reg (rtx x)
2770 rtx new_rtx;
2772 if (REG_P (x)
2773 && HARD_REGISTER_P (x)
2774 && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (x))))
2776 /* Make sure that we generate a REG rather than a CONCAT.
2777 Moves into CONCATs can need nontrivial instructions,
2778 and the whole point of this function is to avoid
2779 using the hard register directly in such a situation. */
2780 generating_concat_p = 0;
2781 new_rtx = gen_reg_rtx (GET_MODE (x));
2782 generating_concat_p = 1;
2783 emit_move_insn (new_rtx, x);
2784 return new_rtx;
2786 return x;
2789 /* Helper function for expand_call.
2790 Return false is EXP is not implementable as a sibling call. */
2792 static bool
2793 can_implement_as_sibling_call_p (tree exp,
2794 rtx structure_value_addr,
2795 tree funtype,
2796 int reg_parm_stack_space ATTRIBUTE_UNUSED,
2797 tree fndecl,
2798 int flags,
2799 tree addr,
2800 const args_size &args_size)
2802 if (!targetm.have_sibcall_epilogue ())
2804 maybe_complain_about_tail_call
2805 (exp,
2806 "machine description does not have"
2807 " a sibcall_epilogue instruction pattern");
2808 return false;
2811 /* Doing sibling call optimization needs some work, since
2812 structure_value_addr can be allocated on the stack.
2813 It does not seem worth the effort since few optimizable
2814 sibling calls will return a structure. */
2815 if (structure_value_addr != NULL_RTX)
2817 maybe_complain_about_tail_call (exp, "callee returns a structure");
2818 return false;
2821 #ifdef REG_PARM_STACK_SPACE
2822 /* If outgoing reg parm stack space changes, we can not do sibcall. */
2823 if (OUTGOING_REG_PARM_STACK_SPACE (funtype)
2824 != OUTGOING_REG_PARM_STACK_SPACE (TREE_TYPE (current_function_decl))
2825 || (reg_parm_stack_space != REG_PARM_STACK_SPACE (current_function_decl)))
2827 maybe_complain_about_tail_call (exp,
2828 "inconsistent size of stack space"
2829 " allocated for arguments which are"
2830 " passed in registers");
2831 return false;
2833 #endif
2835 /* Check whether the target is able to optimize the call
2836 into a sibcall. */
2837 if (!targetm.function_ok_for_sibcall (fndecl, exp))
2839 maybe_complain_about_tail_call (exp,
2840 "target is not able to optimize the"
2841 " call into a sibling call");
2842 return false;
2845 /* Functions that do not return exactly once may not be sibcall
2846 optimized. */
2847 if (flags & ECF_RETURNS_TWICE)
2849 maybe_complain_about_tail_call (exp, "callee returns twice");
2850 return false;
2852 if (flags & ECF_NORETURN)
2854 maybe_complain_about_tail_call (exp, "callee does not return");
2855 return false;
2858 if (TYPE_VOLATILE (TREE_TYPE (TREE_TYPE (addr))))
2860 maybe_complain_about_tail_call (exp, "volatile function type");
2861 return false;
2864 /* If the called function is nested in the current one, it might access
2865 some of the caller's arguments, but could clobber them beforehand if
2866 the argument areas are shared. */
2867 if (fndecl && decl_function_context (fndecl) == current_function_decl)
2869 maybe_complain_about_tail_call (exp, "nested function");
2870 return false;
2873 /* If this function requires more stack slots than the current
2874 function, we cannot change it into a sibling call.
2875 crtl->args.pretend_args_size is not part of the
2876 stack allocated by our caller. */
2877 if (args_size.constant > (crtl->args.size - crtl->args.pretend_args_size))
2879 maybe_complain_about_tail_call (exp,
2880 "callee required more stack slots"
2881 " than the caller");
2882 return false;
2885 /* If the callee pops its own arguments, then it must pop exactly
2886 the same number of arguments as the current function. */
2887 if (targetm.calls.return_pops_args (fndecl, funtype, args_size.constant)
2888 != targetm.calls.return_pops_args (current_function_decl,
2889 TREE_TYPE (current_function_decl),
2890 crtl->args.size))
2892 maybe_complain_about_tail_call (exp,
2893 "inconsistent number of"
2894 " popped arguments");
2895 return false;
2898 if (!lang_hooks.decls.ok_for_sibcall (fndecl))
2900 maybe_complain_about_tail_call (exp, "frontend does not support"
2901 " sibling call");
2902 return false;
2905 /* All checks passed. */
2906 return true;
2909 /* Generate all the code for a CALL_EXPR exp
2910 and return an rtx for its value.
2911 Store the value in TARGET (specified as an rtx) if convenient.
2912 If the value is stored in TARGET then TARGET is returned.
2913 If IGNORE is nonzero, then we ignore the value of the function call. */
2916 expand_call (tree exp, rtx target, int ignore)
2918 /* Nonzero if we are currently expanding a call. */
2919 static int currently_expanding_call = 0;
2921 /* RTX for the function to be called. */
2922 rtx funexp;
2923 /* Sequence of insns to perform a normal "call". */
2924 rtx_insn *normal_call_insns = NULL;
2925 /* Sequence of insns to perform a tail "call". */
2926 rtx_insn *tail_call_insns = NULL;
2927 /* Data type of the function. */
2928 tree funtype;
2929 tree type_arg_types;
2930 tree rettype;
2931 /* Declaration of the function being called,
2932 or 0 if the function is computed (not known by name). */
2933 tree fndecl = 0;
2934 /* The type of the function being called. */
2935 tree fntype;
2936 bool try_tail_call = CALL_EXPR_TAILCALL (exp);
2937 bool must_tail_call = CALL_EXPR_MUST_TAIL_CALL (exp);
2938 int pass;
2940 /* Register in which non-BLKmode value will be returned,
2941 or 0 if no value or if value is BLKmode. */
2942 rtx valreg;
2943 /* Register(s) in which bounds are returned. */
2944 rtx valbnd = NULL;
2945 /* Address where we should return a BLKmode value;
2946 0 if value not BLKmode. */
2947 rtx structure_value_addr = 0;
2948 /* Nonzero if that address is being passed by treating it as
2949 an extra, implicit first parameter. Otherwise,
2950 it is passed by being copied directly into struct_value_rtx. */
2951 int structure_value_addr_parm = 0;
2952 /* Holds the value of implicit argument for the struct value. */
2953 tree structure_value_addr_value = NULL_TREE;
2954 /* Size of aggregate value wanted, or zero if none wanted
2955 or if we are using the non-reentrant PCC calling convention
2956 or expecting the value in registers. */
2957 HOST_WIDE_INT struct_value_size = 0;
2958 /* Nonzero if called function returns an aggregate in memory PCC style,
2959 by returning the address of where to find it. */
2960 int pcc_struct_value = 0;
2961 rtx struct_value = 0;
2963 /* Number of actual parameters in this call, including struct value addr. */
2964 int num_actuals;
2965 /* Number of named args. Args after this are anonymous ones
2966 and they must all go on the stack. */
2967 int n_named_args;
2968 /* Number of complex actual arguments that need to be split. */
2969 int num_complex_actuals = 0;
2971 /* Vector of information about each argument.
2972 Arguments are numbered in the order they will be pushed,
2973 not the order they are written. */
2974 struct arg_data *args;
2976 /* Total size in bytes of all the stack-parms scanned so far. */
2977 struct args_size args_size;
2978 struct args_size adjusted_args_size;
2979 /* Size of arguments before any adjustments (such as rounding). */
2980 int unadjusted_args_size;
2981 /* Data on reg parms scanned so far. */
2982 CUMULATIVE_ARGS args_so_far_v;
2983 cumulative_args_t args_so_far;
2984 /* Nonzero if a reg parm has been scanned. */
2985 int reg_parm_seen;
2986 /* Nonzero if this is an indirect function call. */
2988 /* Nonzero if we must avoid push-insns in the args for this call.
2989 If stack space is allocated for register parameters, but not by the
2990 caller, then it is preallocated in the fixed part of the stack frame.
2991 So the entire argument block must then be preallocated (i.e., we
2992 ignore PUSH_ROUNDING in that case). */
2994 int must_preallocate = !PUSH_ARGS;
2996 /* Size of the stack reserved for parameter registers. */
2997 int reg_parm_stack_space = 0;
2999 /* Address of space preallocated for stack parms
3000 (on machines that lack push insns), or 0 if space not preallocated. */
3001 rtx argblock = 0;
3003 /* Mask of ECF_ and ERF_ flags. */
3004 int flags = 0;
3005 int return_flags = 0;
3006 #ifdef REG_PARM_STACK_SPACE
3007 /* Define the boundary of the register parm stack space that needs to be
3008 saved, if any. */
3009 int low_to_save, high_to_save;
3010 rtx save_area = 0; /* Place that it is saved */
3011 #endif
3013 int initial_highest_arg_in_use = highest_outgoing_arg_in_use;
3014 char *initial_stack_usage_map = stack_usage_map;
3015 char *stack_usage_map_buf = NULL;
3017 int old_stack_allocated;
3019 /* State variables to track stack modifications. */
3020 rtx old_stack_level = 0;
3021 int old_stack_arg_under_construction = 0;
3022 int old_pending_adj = 0;
3023 int old_inhibit_defer_pop = inhibit_defer_pop;
3025 /* Some stack pointer alterations we make are performed via
3026 allocate_dynamic_stack_space. This modifies the stack_pointer_delta,
3027 which we then also need to save/restore along the way. */
3028 int old_stack_pointer_delta = 0;
3030 rtx call_fusage;
3031 tree addr = CALL_EXPR_FN (exp);
3032 int i;
3033 /* The alignment of the stack, in bits. */
3034 unsigned HOST_WIDE_INT preferred_stack_boundary;
3035 /* The alignment of the stack, in bytes. */
3036 unsigned HOST_WIDE_INT preferred_unit_stack_boundary;
3037 /* The static chain value to use for this call. */
3038 rtx static_chain_value;
3039 /* See if this is "nothrow" function call. */
3040 if (TREE_NOTHROW (exp))
3041 flags |= ECF_NOTHROW;
3043 /* See if we can find a DECL-node for the actual function, and get the
3044 function attributes (flags) from the function decl or type node. */
3045 fndecl = get_callee_fndecl (exp);
3046 if (fndecl)
3048 fntype = TREE_TYPE (fndecl);
3049 flags |= flags_from_decl_or_type (fndecl);
3050 return_flags |= decl_return_flags (fndecl);
3052 else
3054 fntype = TREE_TYPE (TREE_TYPE (addr));
3055 flags |= flags_from_decl_or_type (fntype);
3056 if (CALL_EXPR_BY_DESCRIPTOR (exp))
3057 flags |= ECF_BY_DESCRIPTOR;
3059 rettype = TREE_TYPE (exp);
3061 struct_value = targetm.calls.struct_value_rtx (fntype, 0);
3063 /* Warn if this value is an aggregate type,
3064 regardless of which calling convention we are using for it. */
3065 if (AGGREGATE_TYPE_P (rettype))
3066 warning (OPT_Waggregate_return, "function call has aggregate value");
3068 /* If the result of a non looping pure or const function call is
3069 ignored (or void), and none of its arguments are volatile, we can
3070 avoid expanding the call and just evaluate the arguments for
3071 side-effects. */
3072 if ((flags & (ECF_CONST | ECF_PURE))
3073 && (!(flags & ECF_LOOPING_CONST_OR_PURE))
3074 && (ignore || target == const0_rtx
3075 || TYPE_MODE (rettype) == VOIDmode))
3077 bool volatilep = false;
3078 tree arg;
3079 call_expr_arg_iterator iter;
3081 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
3082 if (TREE_THIS_VOLATILE (arg))
3084 volatilep = true;
3085 break;
3088 if (! volatilep)
3090 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
3091 expand_expr (arg, const0_rtx, VOIDmode, EXPAND_NORMAL);
3092 return const0_rtx;
3096 #ifdef REG_PARM_STACK_SPACE
3097 reg_parm_stack_space = REG_PARM_STACK_SPACE (!fndecl ? fntype : fndecl);
3098 #endif
3100 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))
3101 && reg_parm_stack_space > 0 && PUSH_ARGS)
3102 must_preallocate = 1;
3104 /* Set up a place to return a structure. */
3106 /* Cater to broken compilers. */
3107 if (aggregate_value_p (exp, fntype))
3109 /* This call returns a big structure. */
3110 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
3112 #ifdef PCC_STATIC_STRUCT_RETURN
3114 pcc_struct_value = 1;
3116 #else /* not PCC_STATIC_STRUCT_RETURN */
3118 struct_value_size = int_size_in_bytes (rettype);
3120 /* Even if it is semantically safe to use the target as the return
3121 slot, it may be not sufficiently aligned for the return type. */
3122 if (CALL_EXPR_RETURN_SLOT_OPT (exp)
3123 && target
3124 && MEM_P (target)
3125 && !(MEM_ALIGN (target) < TYPE_ALIGN (rettype)
3126 && SLOW_UNALIGNED_ACCESS (TYPE_MODE (rettype),
3127 MEM_ALIGN (target))))
3128 structure_value_addr = XEXP (target, 0);
3129 else
3131 /* For variable-sized objects, we must be called with a target
3132 specified. If we were to allocate space on the stack here,
3133 we would have no way of knowing when to free it. */
3134 rtx d = assign_temp (rettype, 1, 1);
3135 structure_value_addr = XEXP (d, 0);
3136 target = 0;
3139 #endif /* not PCC_STATIC_STRUCT_RETURN */
3142 /* Figure out the amount to which the stack should be aligned. */
3143 preferred_stack_boundary = PREFERRED_STACK_BOUNDARY;
3144 if (fndecl)
3146 struct cgraph_rtl_info *i = cgraph_node::rtl_info (fndecl);
3147 /* Without automatic stack alignment, we can't increase preferred
3148 stack boundary. With automatic stack alignment, it is
3149 unnecessary since unless we can guarantee that all callers will
3150 align the outgoing stack properly, callee has to align its
3151 stack anyway. */
3152 if (i
3153 && i->preferred_incoming_stack_boundary
3154 && i->preferred_incoming_stack_boundary < preferred_stack_boundary)
3155 preferred_stack_boundary = i->preferred_incoming_stack_boundary;
3158 /* Operand 0 is a pointer-to-function; get the type of the function. */
3159 funtype = TREE_TYPE (addr);
3160 gcc_assert (POINTER_TYPE_P (funtype));
3161 funtype = TREE_TYPE (funtype);
3163 /* Count whether there are actual complex arguments that need to be split
3164 into their real and imaginary parts. Munge the type_arg_types
3165 appropriately here as well. */
3166 if (targetm.calls.split_complex_arg)
3168 call_expr_arg_iterator iter;
3169 tree arg;
3170 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
3172 tree type = TREE_TYPE (arg);
3173 if (type && TREE_CODE (type) == COMPLEX_TYPE
3174 && targetm.calls.split_complex_arg (type))
3175 num_complex_actuals++;
3177 type_arg_types = split_complex_types (TYPE_ARG_TYPES (funtype));
3179 else
3180 type_arg_types = TYPE_ARG_TYPES (funtype);
3182 if (flags & ECF_MAY_BE_ALLOCA)
3183 cfun->calls_alloca = 1;
3185 /* If struct_value_rtx is 0, it means pass the address
3186 as if it were an extra parameter. Put the argument expression
3187 in structure_value_addr_value. */
3188 if (structure_value_addr && struct_value == 0)
3190 /* If structure_value_addr is a REG other than
3191 virtual_outgoing_args_rtx, we can use always use it. If it
3192 is not a REG, we must always copy it into a register.
3193 If it is virtual_outgoing_args_rtx, we must copy it to another
3194 register in some cases. */
3195 rtx temp = (!REG_P (structure_value_addr)
3196 || (ACCUMULATE_OUTGOING_ARGS
3197 && stack_arg_under_construction
3198 && structure_value_addr == virtual_outgoing_args_rtx)
3199 ? copy_addr_to_reg (convert_memory_address
3200 (Pmode, structure_value_addr))
3201 : structure_value_addr);
3203 structure_value_addr_value =
3204 make_tree (build_pointer_type (TREE_TYPE (funtype)), temp);
3205 structure_value_addr_parm = CALL_WITH_BOUNDS_P (exp) ? 2 : 1;
3208 /* Count the arguments and set NUM_ACTUALS. */
3209 num_actuals =
3210 call_expr_nargs (exp) + num_complex_actuals + structure_value_addr_parm;
3212 /* Compute number of named args.
3213 First, do a raw count of the args for INIT_CUMULATIVE_ARGS. */
3215 if (type_arg_types != 0)
3216 n_named_args
3217 = (list_length (type_arg_types)
3218 /* Count the struct value address, if it is passed as a parm. */
3219 + structure_value_addr_parm);
3220 else
3221 /* If we know nothing, treat all args as named. */
3222 n_named_args = num_actuals;
3224 /* Start updating where the next arg would go.
3226 On some machines (such as the PA) indirect calls have a different
3227 calling convention than normal calls. The fourth argument in
3228 INIT_CUMULATIVE_ARGS tells the backend if this is an indirect call
3229 or not. */
3230 INIT_CUMULATIVE_ARGS (args_so_far_v, funtype, NULL_RTX, fndecl, n_named_args);
3231 args_so_far = pack_cumulative_args (&args_so_far_v);
3233 /* Now possibly adjust the number of named args.
3234 Normally, don't include the last named arg if anonymous args follow.
3235 We do include the last named arg if
3236 targetm.calls.strict_argument_naming() returns nonzero.
3237 (If no anonymous args follow, the result of list_length is actually
3238 one too large. This is harmless.)
3240 If targetm.calls.pretend_outgoing_varargs_named() returns
3241 nonzero, and targetm.calls.strict_argument_naming() returns zero,
3242 this machine will be able to place unnamed args that were passed
3243 in registers into the stack. So treat all args as named. This
3244 allows the insns emitting for a specific argument list to be
3245 independent of the function declaration.
3247 If targetm.calls.pretend_outgoing_varargs_named() returns zero,
3248 we do not have any reliable way to pass unnamed args in
3249 registers, so we must force them into memory. */
3251 if (type_arg_types != 0
3252 && targetm.calls.strict_argument_naming (args_so_far))
3254 else if (type_arg_types != 0
3255 && ! targetm.calls.pretend_outgoing_varargs_named (args_so_far))
3256 /* Don't include the last named arg. */
3257 --n_named_args;
3258 else
3259 /* Treat all args as named. */
3260 n_named_args = num_actuals;
3262 /* Make a vector to hold all the information about each arg. */
3263 args = XCNEWVEC (struct arg_data, num_actuals);
3265 /* Build up entries in the ARGS array, compute the size of the
3266 arguments into ARGS_SIZE, etc. */
3267 initialize_argument_information (num_actuals, args, &args_size,
3268 n_named_args, exp,
3269 structure_value_addr_value, fndecl, fntype,
3270 args_so_far, reg_parm_stack_space,
3271 &old_stack_level, &old_pending_adj,
3272 &must_preallocate, &flags,
3273 &try_tail_call, CALL_FROM_THUNK_P (exp));
3275 if (args_size.var)
3276 must_preallocate = 1;
3278 /* Now make final decision about preallocating stack space. */
3279 must_preallocate = finalize_must_preallocate (must_preallocate,
3280 num_actuals, args,
3281 &args_size);
3283 /* If the structure value address will reference the stack pointer, we
3284 must stabilize it. We don't need to do this if we know that we are
3285 not going to adjust the stack pointer in processing this call. */
3287 if (structure_value_addr
3288 && (reg_mentioned_p (virtual_stack_dynamic_rtx, structure_value_addr)
3289 || reg_mentioned_p (virtual_outgoing_args_rtx,
3290 structure_value_addr))
3291 && (args_size.var
3292 || (!ACCUMULATE_OUTGOING_ARGS && args_size.constant)))
3293 structure_value_addr = copy_to_reg (structure_value_addr);
3295 /* Tail calls can make things harder to debug, and we've traditionally
3296 pushed these optimizations into -O2. Don't try if we're already
3297 expanding a call, as that means we're an argument. Don't try if
3298 there's cleanups, as we know there's code to follow the call. */
3300 if (currently_expanding_call++ != 0
3301 || !flag_optimize_sibling_calls
3302 || args_size.var
3303 || dbg_cnt (tail_call) == false)
3304 try_tail_call = 0;
3306 /* If the user has marked the function as requiring tail-call
3307 optimization, attempt it. */
3308 if (must_tail_call)
3309 try_tail_call = 1;
3311 /* Rest of purposes for tail call optimizations to fail. */
3312 if (try_tail_call)
3313 try_tail_call = can_implement_as_sibling_call_p (exp,
3314 structure_value_addr,
3315 funtype,
3316 reg_parm_stack_space,
3317 fndecl,
3318 flags, addr, args_size);
3320 /* Check if caller and callee disagree in promotion of function
3321 return value. */
3322 if (try_tail_call)
3324 machine_mode caller_mode, caller_promoted_mode;
3325 machine_mode callee_mode, callee_promoted_mode;
3326 int caller_unsignedp, callee_unsignedp;
3327 tree caller_res = DECL_RESULT (current_function_decl);
3329 caller_unsignedp = TYPE_UNSIGNED (TREE_TYPE (caller_res));
3330 caller_mode = DECL_MODE (caller_res);
3331 callee_unsignedp = TYPE_UNSIGNED (TREE_TYPE (funtype));
3332 callee_mode = TYPE_MODE (TREE_TYPE (funtype));
3333 caller_promoted_mode
3334 = promote_function_mode (TREE_TYPE (caller_res), caller_mode,
3335 &caller_unsignedp,
3336 TREE_TYPE (current_function_decl), 1);
3337 callee_promoted_mode
3338 = promote_function_mode (TREE_TYPE (funtype), callee_mode,
3339 &callee_unsignedp,
3340 funtype, 1);
3341 if (caller_mode != VOIDmode
3342 && (caller_promoted_mode != callee_promoted_mode
3343 || ((caller_mode != caller_promoted_mode
3344 || callee_mode != callee_promoted_mode)
3345 && (caller_unsignedp != callee_unsignedp
3346 || GET_MODE_BITSIZE (caller_mode)
3347 < GET_MODE_BITSIZE (callee_mode)))))
3349 try_tail_call = 0;
3350 maybe_complain_about_tail_call (exp,
3351 "caller and callee disagree in"
3352 " promotion of function"
3353 " return value");
3357 /* Ensure current function's preferred stack boundary is at least
3358 what we need. Stack alignment may also increase preferred stack
3359 boundary. */
3360 if (crtl->preferred_stack_boundary < preferred_stack_boundary)
3361 crtl->preferred_stack_boundary = preferred_stack_boundary;
3362 else
3363 preferred_stack_boundary = crtl->preferred_stack_boundary;
3365 preferred_unit_stack_boundary = preferred_stack_boundary / BITS_PER_UNIT;
3367 /* We want to make two insn chains; one for a sibling call, the other
3368 for a normal call. We will select one of the two chains after
3369 initial RTL generation is complete. */
3370 for (pass = try_tail_call ? 0 : 1; pass < 2; pass++)
3372 int sibcall_failure = 0;
3373 /* We want to emit any pending stack adjustments before the tail
3374 recursion "call". That way we know any adjustment after the tail
3375 recursion call can be ignored if we indeed use the tail
3376 call expansion. */
3377 saved_pending_stack_adjust save;
3378 rtx_insn *insns, *before_call, *after_args;
3379 rtx next_arg_reg;
3381 if (pass == 0)
3383 /* State variables we need to save and restore between
3384 iterations. */
3385 save_pending_stack_adjust (&save);
3387 if (pass)
3388 flags &= ~ECF_SIBCALL;
3389 else
3390 flags |= ECF_SIBCALL;
3392 /* Other state variables that we must reinitialize each time
3393 through the loop (that are not initialized by the loop itself). */
3394 argblock = 0;
3395 call_fusage = 0;
3397 /* Start a new sequence for the normal call case.
3399 From this point on, if the sibling call fails, we want to set
3400 sibcall_failure instead of continuing the loop. */
3401 start_sequence ();
3403 /* Don't let pending stack adjusts add up to too much.
3404 Also, do all pending adjustments now if there is any chance
3405 this might be a call to alloca or if we are expanding a sibling
3406 call sequence.
3407 Also do the adjustments before a throwing call, otherwise
3408 exception handling can fail; PR 19225. */
3409 if (pending_stack_adjust >= 32
3410 || (pending_stack_adjust > 0
3411 && (flags & ECF_MAY_BE_ALLOCA))
3412 || (pending_stack_adjust > 0
3413 && flag_exceptions && !(flags & ECF_NOTHROW))
3414 || pass == 0)
3415 do_pending_stack_adjust ();
3417 /* Precompute any arguments as needed. */
3418 if (pass)
3419 precompute_arguments (num_actuals, args);
3421 /* Now we are about to start emitting insns that can be deleted
3422 if a libcall is deleted. */
3423 if (pass && (flags & ECF_MALLOC))
3424 start_sequence ();
3426 if (pass == 0
3427 && crtl->stack_protect_guard
3428 && targetm.stack_protect_runtime_enabled_p ())
3429 stack_protect_epilogue ();
3431 adjusted_args_size = args_size;
3432 /* Compute the actual size of the argument block required. The variable
3433 and constant sizes must be combined, the size may have to be rounded,
3434 and there may be a minimum required size. When generating a sibcall
3435 pattern, do not round up, since we'll be re-using whatever space our
3436 caller provided. */
3437 unadjusted_args_size
3438 = compute_argument_block_size (reg_parm_stack_space,
3439 &adjusted_args_size,
3440 fndecl, fntype,
3441 (pass == 0 ? 0
3442 : preferred_stack_boundary));
3444 old_stack_allocated = stack_pointer_delta - pending_stack_adjust;
3446 /* The argument block when performing a sibling call is the
3447 incoming argument block. */
3448 if (pass == 0)
3450 argblock = crtl->args.internal_arg_pointer;
3451 if (STACK_GROWS_DOWNWARD)
3452 argblock
3453 = plus_constant (Pmode, argblock, crtl->args.pretend_args_size);
3454 else
3455 argblock
3456 = plus_constant (Pmode, argblock, -crtl->args.pretend_args_size);
3458 stored_args_map = sbitmap_alloc (args_size.constant);
3459 bitmap_clear (stored_args_map);
3462 /* If we have no actual push instructions, or shouldn't use them,
3463 make space for all args right now. */
3464 else if (adjusted_args_size.var != 0)
3466 if (old_stack_level == 0)
3468 emit_stack_save (SAVE_BLOCK, &old_stack_level);
3469 old_stack_pointer_delta = stack_pointer_delta;
3470 old_pending_adj = pending_stack_adjust;
3471 pending_stack_adjust = 0;
3472 /* stack_arg_under_construction says whether a stack arg is
3473 being constructed at the old stack level. Pushing the stack
3474 gets a clean outgoing argument block. */
3475 old_stack_arg_under_construction = stack_arg_under_construction;
3476 stack_arg_under_construction = 0;
3478 argblock = push_block (ARGS_SIZE_RTX (adjusted_args_size), 0, 0);
3479 if (flag_stack_usage_info)
3480 current_function_has_unbounded_dynamic_stack_size = 1;
3482 else
3484 /* Note that we must go through the motions of allocating an argument
3485 block even if the size is zero because we may be storing args
3486 in the area reserved for register arguments, which may be part of
3487 the stack frame. */
3489 int needed = adjusted_args_size.constant;
3491 /* Store the maximum argument space used. It will be pushed by
3492 the prologue (if ACCUMULATE_OUTGOING_ARGS, or stack overflow
3493 checking). */
3495 if (needed > crtl->outgoing_args_size)
3496 crtl->outgoing_args_size = needed;
3498 if (must_preallocate)
3500 if (ACCUMULATE_OUTGOING_ARGS)
3502 /* Since the stack pointer will never be pushed, it is
3503 possible for the evaluation of a parm to clobber
3504 something we have already written to the stack.
3505 Since most function calls on RISC machines do not use
3506 the stack, this is uncommon, but must work correctly.
3508 Therefore, we save any area of the stack that was already
3509 written and that we are using. Here we set up to do this
3510 by making a new stack usage map from the old one. The
3511 actual save will be done by store_one_arg.
3513 Another approach might be to try to reorder the argument
3514 evaluations to avoid this conflicting stack usage. */
3516 /* Since we will be writing into the entire argument area,
3517 the map must be allocated for its entire size, not just
3518 the part that is the responsibility of the caller. */
3519 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
3520 needed += reg_parm_stack_space;
3522 if (ARGS_GROW_DOWNWARD)
3523 highest_outgoing_arg_in_use
3524 = MAX (initial_highest_arg_in_use, needed + 1);
3525 else
3526 highest_outgoing_arg_in_use
3527 = MAX (initial_highest_arg_in_use, needed);
3529 free (stack_usage_map_buf);
3530 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use);
3531 stack_usage_map = stack_usage_map_buf;
3533 if (initial_highest_arg_in_use)
3534 memcpy (stack_usage_map, initial_stack_usage_map,
3535 initial_highest_arg_in_use);
3537 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use)
3538 memset (&stack_usage_map[initial_highest_arg_in_use], 0,
3539 (highest_outgoing_arg_in_use
3540 - initial_highest_arg_in_use));
3541 needed = 0;
3543 /* The address of the outgoing argument list must not be
3544 copied to a register here, because argblock would be left
3545 pointing to the wrong place after the call to
3546 allocate_dynamic_stack_space below. */
3548 argblock = virtual_outgoing_args_rtx;
3550 else
3552 if (inhibit_defer_pop == 0)
3554 /* Try to reuse some or all of the pending_stack_adjust
3555 to get this space. */
3556 needed
3557 = (combine_pending_stack_adjustment_and_call
3558 (unadjusted_args_size,
3559 &adjusted_args_size,
3560 preferred_unit_stack_boundary));
3562 /* combine_pending_stack_adjustment_and_call computes
3563 an adjustment before the arguments are allocated.
3564 Account for them and see whether or not the stack
3565 needs to go up or down. */
3566 needed = unadjusted_args_size - needed;
3568 if (needed < 0)
3570 /* We're releasing stack space. */
3571 /* ??? We can avoid any adjustment at all if we're
3572 already aligned. FIXME. */
3573 pending_stack_adjust = -needed;
3574 do_pending_stack_adjust ();
3575 needed = 0;
3577 else
3578 /* We need to allocate space. We'll do that in
3579 push_block below. */
3580 pending_stack_adjust = 0;
3583 /* Special case this because overhead of `push_block' in
3584 this case is non-trivial. */
3585 if (needed == 0)
3586 argblock = virtual_outgoing_args_rtx;
3587 else
3589 argblock = push_block (GEN_INT (needed), 0, 0);
3590 if (ARGS_GROW_DOWNWARD)
3591 argblock = plus_constant (Pmode, argblock, needed);
3594 /* We only really need to call `copy_to_reg' in the case
3595 where push insns are going to be used to pass ARGBLOCK
3596 to a function call in ARGS. In that case, the stack
3597 pointer changes value from the allocation point to the
3598 call point, and hence the value of
3599 VIRTUAL_OUTGOING_ARGS_RTX changes as well. But might
3600 as well always do it. */
3601 argblock = copy_to_reg (argblock);
3606 if (ACCUMULATE_OUTGOING_ARGS)
3608 /* The save/restore code in store_one_arg handles all
3609 cases except one: a constructor call (including a C
3610 function returning a BLKmode struct) to initialize
3611 an argument. */
3612 if (stack_arg_under_construction)
3614 rtx push_size
3615 = GEN_INT (adjusted_args_size.constant
3616 + (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype
3617 : TREE_TYPE (fndecl))) ? 0
3618 : reg_parm_stack_space));
3619 if (old_stack_level == 0)
3621 emit_stack_save (SAVE_BLOCK, &old_stack_level);
3622 old_stack_pointer_delta = stack_pointer_delta;
3623 old_pending_adj = pending_stack_adjust;
3624 pending_stack_adjust = 0;
3625 /* stack_arg_under_construction says whether a stack
3626 arg is being constructed at the old stack level.
3627 Pushing the stack gets a clean outgoing argument
3628 block. */
3629 old_stack_arg_under_construction
3630 = stack_arg_under_construction;
3631 stack_arg_under_construction = 0;
3632 /* Make a new map for the new argument list. */
3633 free (stack_usage_map_buf);
3634 stack_usage_map_buf = XCNEWVEC (char, highest_outgoing_arg_in_use);
3635 stack_usage_map = stack_usage_map_buf;
3636 highest_outgoing_arg_in_use = 0;
3638 /* We can pass TRUE as the 4th argument because we just
3639 saved the stack pointer and will restore it right after
3640 the call. */
3641 allocate_dynamic_stack_space (push_size, 0,
3642 BIGGEST_ALIGNMENT, true);
3645 /* If argument evaluation might modify the stack pointer,
3646 copy the address of the argument list to a register. */
3647 for (i = 0; i < num_actuals; i++)
3648 if (args[i].pass_on_stack)
3650 argblock = copy_addr_to_reg (argblock);
3651 break;
3655 compute_argument_addresses (args, argblock, num_actuals);
3657 /* Stack is properly aligned, pops can't safely be deferred during
3658 the evaluation of the arguments. */
3659 NO_DEFER_POP;
3661 /* Precompute all register parameters. It isn't safe to compute
3662 anything once we have started filling any specific hard regs.
3663 TLS symbols sometimes need a call to resolve. Precompute
3664 register parameters before any stack pointer manipulation
3665 to avoid unaligned stack in the called function. */
3666 precompute_register_parameters (num_actuals, args, &reg_parm_seen);
3668 OK_DEFER_POP;
3670 /* Perform stack alignment before the first push (the last arg). */
3671 if (argblock == 0
3672 && adjusted_args_size.constant > reg_parm_stack_space
3673 && adjusted_args_size.constant != unadjusted_args_size)
3675 /* When the stack adjustment is pending, we get better code
3676 by combining the adjustments. */
3677 if (pending_stack_adjust
3678 && ! inhibit_defer_pop)
3680 pending_stack_adjust
3681 = (combine_pending_stack_adjustment_and_call
3682 (unadjusted_args_size,
3683 &adjusted_args_size,
3684 preferred_unit_stack_boundary));
3685 do_pending_stack_adjust ();
3687 else if (argblock == 0)
3688 anti_adjust_stack (GEN_INT (adjusted_args_size.constant
3689 - unadjusted_args_size));
3691 /* Now that the stack is properly aligned, pops can't safely
3692 be deferred during the evaluation of the arguments. */
3693 NO_DEFER_POP;
3695 /* Record the maximum pushed stack space size. We need to delay
3696 doing it this far to take into account the optimization done
3697 by combine_pending_stack_adjustment_and_call. */
3698 if (flag_stack_usage_info
3699 && !ACCUMULATE_OUTGOING_ARGS
3700 && pass
3701 && adjusted_args_size.var == 0)
3703 int pushed = adjusted_args_size.constant + pending_stack_adjust;
3704 if (pushed > current_function_pushed_stack_size)
3705 current_function_pushed_stack_size = pushed;
3708 funexp = rtx_for_function_call (fndecl, addr);
3710 if (CALL_EXPR_STATIC_CHAIN (exp))
3711 static_chain_value = expand_normal (CALL_EXPR_STATIC_CHAIN (exp));
3712 else
3713 static_chain_value = 0;
3715 #ifdef REG_PARM_STACK_SPACE
3716 /* Save the fixed argument area if it's part of the caller's frame and
3717 is clobbered by argument setup for this call. */
3718 if (ACCUMULATE_OUTGOING_ARGS && pass)
3719 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock,
3720 &low_to_save, &high_to_save);
3721 #endif
3723 /* Now store (and compute if necessary) all non-register parms.
3724 These come before register parms, since they can require block-moves,
3725 which could clobber the registers used for register parms.
3726 Parms which have partial registers are not stored here,
3727 but we do preallocate space here if they want that. */
3729 for (i = 0; i < num_actuals; i++)
3731 /* Delay bounds until all other args are stored. */
3732 if (POINTER_BOUNDS_P (args[i].tree_value))
3733 continue;
3734 else if (args[i].reg == 0 || args[i].pass_on_stack)
3736 rtx_insn *before_arg = get_last_insn ();
3738 /* We don't allow passing huge (> 2^30 B) arguments
3739 by value. It would cause an overflow later on. */
3740 if (adjusted_args_size.constant
3741 >= (1 << (HOST_BITS_PER_INT - 2)))
3743 sorry ("passing too large argument on stack");
3744 continue;
3747 if (store_one_arg (&args[i], argblock, flags,
3748 adjusted_args_size.var != 0,
3749 reg_parm_stack_space)
3750 || (pass == 0
3751 && check_sibcall_argument_overlap (before_arg,
3752 &args[i], 1)))
3753 sibcall_failure = 1;
3756 if (args[i].stack)
3757 call_fusage
3758 = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[i].tree_value)),
3759 gen_rtx_USE (VOIDmode, args[i].stack),
3760 call_fusage);
3763 /* If we have a parm that is passed in registers but not in memory
3764 and whose alignment does not permit a direct copy into registers,
3765 make a group of pseudos that correspond to each register that we
3766 will later fill. */
3767 if (STRICT_ALIGNMENT)
3768 store_unaligned_arguments_into_pseudos (args, num_actuals);
3770 /* Now store any partially-in-registers parm.
3771 This is the last place a block-move can happen. */
3772 if (reg_parm_seen)
3773 for (i = 0; i < num_actuals; i++)
3774 if (args[i].partial != 0 && ! args[i].pass_on_stack)
3776 rtx_insn *before_arg = get_last_insn ();
3778 /* On targets with weird calling conventions (e.g. PA) it's
3779 hard to ensure that all cases of argument overlap between
3780 stack and registers work. Play it safe and bail out. */
3781 if (ARGS_GROW_DOWNWARD && !STACK_GROWS_DOWNWARD)
3783 sibcall_failure = 1;
3784 break;
3787 if (store_one_arg (&args[i], argblock, flags,
3788 adjusted_args_size.var != 0,
3789 reg_parm_stack_space)
3790 || (pass == 0
3791 && check_sibcall_argument_overlap (before_arg,
3792 &args[i], 1)))
3793 sibcall_failure = 1;
3796 bool any_regs = false;
3797 for (i = 0; i < num_actuals; i++)
3798 if (args[i].reg != NULL_RTX)
3800 any_regs = true;
3801 targetm.calls.call_args (args[i].reg, funtype);
3803 if (!any_regs)
3804 targetm.calls.call_args (pc_rtx, funtype);
3806 /* Figure out the register where the value, if any, will come back. */
3807 valreg = 0;
3808 valbnd = 0;
3809 if (TYPE_MODE (rettype) != VOIDmode
3810 && ! structure_value_addr)
3812 if (pcc_struct_value)
3814 valreg = hard_function_value (build_pointer_type (rettype),
3815 fndecl, NULL, (pass == 0));
3816 if (CALL_WITH_BOUNDS_P (exp))
3817 valbnd = targetm.calls.
3818 chkp_function_value_bounds (build_pointer_type (rettype),
3819 fndecl, (pass == 0));
3821 else
3823 valreg = hard_function_value (rettype, fndecl, fntype,
3824 (pass == 0));
3825 if (CALL_WITH_BOUNDS_P (exp))
3826 valbnd = targetm.calls.chkp_function_value_bounds (rettype,
3827 fndecl,
3828 (pass == 0));
3831 /* If VALREG is a PARALLEL whose first member has a zero
3832 offset, use that. This is for targets such as m68k that
3833 return the same value in multiple places. */
3834 if (GET_CODE (valreg) == PARALLEL)
3836 rtx elem = XVECEXP (valreg, 0, 0);
3837 rtx where = XEXP (elem, 0);
3838 rtx offset = XEXP (elem, 1);
3839 if (offset == const0_rtx
3840 && GET_MODE (where) == GET_MODE (valreg))
3841 valreg = where;
3845 /* Store all bounds not passed in registers. */
3846 for (i = 0; i < num_actuals; i++)
3848 if (POINTER_BOUNDS_P (args[i].tree_value)
3849 && !args[i].reg)
3850 store_bounds (&args[i],
3851 args[i].pointer_arg == -1
3852 ? NULL
3853 : &args[args[i].pointer_arg]);
3856 /* If register arguments require space on the stack and stack space
3857 was not preallocated, allocate stack space here for arguments
3858 passed in registers. */
3859 if (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))
3860 && !ACCUMULATE_OUTGOING_ARGS
3861 && must_preallocate == 0 && reg_parm_stack_space > 0)
3862 anti_adjust_stack (GEN_INT (reg_parm_stack_space));
3864 /* Pass the function the address in which to return a
3865 structure value. */
3866 if (pass != 0 && structure_value_addr && ! structure_value_addr_parm)
3868 structure_value_addr
3869 = convert_memory_address (Pmode, structure_value_addr);
3870 emit_move_insn (struct_value,
3871 force_reg (Pmode,
3872 force_operand (structure_value_addr,
3873 NULL_RTX)));
3875 if (REG_P (struct_value))
3876 use_reg (&call_fusage, struct_value);
3879 after_args = get_last_insn ();
3880 funexp = prepare_call_address (fndecl ? fndecl : fntype, funexp,
3881 static_chain_value, &call_fusage,
3882 reg_parm_seen, flags);
3884 load_register_parameters (args, num_actuals, &call_fusage, flags,
3885 pass == 0, &sibcall_failure);
3887 /* Save a pointer to the last insn before the call, so that we can
3888 later safely search backwards to find the CALL_INSN. */
3889 before_call = get_last_insn ();
3891 /* Set up next argument register. For sibling calls on machines
3892 with register windows this should be the incoming register. */
3893 if (pass == 0)
3894 next_arg_reg = targetm.calls.function_incoming_arg (args_so_far,
3895 VOIDmode,
3896 void_type_node,
3897 true);
3898 else
3899 next_arg_reg = targetm.calls.function_arg (args_so_far,
3900 VOIDmode, void_type_node,
3901 true);
3903 if (pass == 1 && (return_flags & ERF_RETURNS_ARG))
3905 int arg_nr = return_flags & ERF_RETURN_ARG_MASK;
3906 arg_nr = num_actuals - arg_nr - 1;
3907 if (arg_nr >= 0
3908 && arg_nr < num_actuals
3909 && args[arg_nr].reg
3910 && valreg
3911 && REG_P (valreg)
3912 && GET_MODE (args[arg_nr].reg) == GET_MODE (valreg))
3913 call_fusage
3914 = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[arg_nr].tree_value)),
3915 gen_rtx_SET (valreg, args[arg_nr].reg),
3916 call_fusage);
3918 /* All arguments and registers used for the call must be set up by
3919 now! */
3921 /* Stack must be properly aligned now. */
3922 gcc_assert (!pass
3923 || !(stack_pointer_delta % preferred_unit_stack_boundary));
3925 /* Generate the actual call instruction. */
3926 emit_call_1 (funexp, exp, fndecl, funtype, unadjusted_args_size,
3927 adjusted_args_size.constant, struct_value_size,
3928 next_arg_reg, valreg, old_inhibit_defer_pop, call_fusage,
3929 flags, args_so_far);
3931 if (flag_ipa_ra)
3933 rtx_call_insn *last;
3934 rtx datum = NULL_RTX;
3935 if (fndecl != NULL_TREE)
3937 datum = XEXP (DECL_RTL (fndecl), 0);
3938 gcc_assert (datum != NULL_RTX
3939 && GET_CODE (datum) == SYMBOL_REF);
3941 last = last_call_insn ();
3942 add_reg_note (last, REG_CALL_DECL, datum);
3945 /* If the call setup or the call itself overlaps with anything
3946 of the argument setup we probably clobbered our call address.
3947 In that case we can't do sibcalls. */
3948 if (pass == 0
3949 && check_sibcall_argument_overlap (after_args, 0, 0))
3950 sibcall_failure = 1;
3952 /* If a non-BLKmode value is returned at the most significant end
3953 of a register, shift the register right by the appropriate amount
3954 and update VALREG accordingly. BLKmode values are handled by the
3955 group load/store machinery below. */
3956 if (!structure_value_addr
3957 && !pcc_struct_value
3958 && TYPE_MODE (rettype) != VOIDmode
3959 && TYPE_MODE (rettype) != BLKmode
3960 && REG_P (valreg)
3961 && targetm.calls.return_in_msb (rettype))
3963 if (shift_return_value (TYPE_MODE (rettype), false, valreg))
3964 sibcall_failure = 1;
3965 valreg = gen_rtx_REG (TYPE_MODE (rettype), REGNO (valreg));
3968 if (pass && (flags & ECF_MALLOC))
3970 rtx temp = gen_reg_rtx (GET_MODE (valreg));
3971 rtx_insn *last, *insns;
3973 /* The return value from a malloc-like function is a pointer. */
3974 if (TREE_CODE (rettype) == POINTER_TYPE)
3975 mark_reg_pointer (temp, MALLOC_ABI_ALIGNMENT);
3977 emit_move_insn (temp, valreg);
3979 /* The return value from a malloc-like function can not alias
3980 anything else. */
3981 last = get_last_insn ();
3982 add_reg_note (last, REG_NOALIAS, temp);
3984 /* Write out the sequence. */
3985 insns = get_insns ();
3986 end_sequence ();
3987 emit_insn (insns);
3988 valreg = temp;
3991 /* For calls to `setjmp', etc., inform
3992 function.c:setjmp_warnings that it should complain if
3993 nonvolatile values are live. For functions that cannot
3994 return, inform flow that control does not fall through. */
3996 if ((flags & ECF_NORETURN) || pass == 0)
3998 /* The barrier must be emitted
3999 immediately after the CALL_INSN. Some ports emit more
4000 than just a CALL_INSN above, so we must search for it here. */
4002 rtx_insn *last = get_last_insn ();
4003 while (!CALL_P (last))
4005 last = PREV_INSN (last);
4006 /* There was no CALL_INSN? */
4007 gcc_assert (last != before_call);
4010 emit_barrier_after (last);
4012 /* Stack adjustments after a noreturn call are dead code.
4013 However when NO_DEFER_POP is in effect, we must preserve
4014 stack_pointer_delta. */
4015 if (inhibit_defer_pop == 0)
4017 stack_pointer_delta = old_stack_allocated;
4018 pending_stack_adjust = 0;
4022 /* If value type not void, return an rtx for the value. */
4024 if (TYPE_MODE (rettype) == VOIDmode
4025 || ignore)
4026 target = const0_rtx;
4027 else if (structure_value_addr)
4029 if (target == 0 || !MEM_P (target))
4031 target
4032 = gen_rtx_MEM (TYPE_MODE (rettype),
4033 memory_address (TYPE_MODE (rettype),
4034 structure_value_addr));
4035 set_mem_attributes (target, rettype, 1);
4038 else if (pcc_struct_value)
4040 /* This is the special C++ case where we need to
4041 know what the true target was. We take care to
4042 never use this value more than once in one expression. */
4043 target = gen_rtx_MEM (TYPE_MODE (rettype),
4044 copy_to_reg (valreg));
4045 set_mem_attributes (target, rettype, 1);
4047 /* Handle calls that return values in multiple non-contiguous locations.
4048 The Irix 6 ABI has examples of this. */
4049 else if (GET_CODE (valreg) == PARALLEL)
4051 if (target == 0)
4052 target = emit_group_move_into_temps (valreg);
4053 else if (rtx_equal_p (target, valreg))
4055 else if (GET_CODE (target) == PARALLEL)
4056 /* Handle the result of a emit_group_move_into_temps
4057 call in the previous pass. */
4058 emit_group_move (target, valreg);
4059 else
4060 emit_group_store (target, valreg, rettype,
4061 int_size_in_bytes (rettype));
4063 else if (target
4064 && GET_MODE (target) == TYPE_MODE (rettype)
4065 && GET_MODE (target) == GET_MODE (valreg))
4067 bool may_overlap = false;
4069 /* We have to copy a return value in a CLASS_LIKELY_SPILLED hard
4070 reg to a plain register. */
4071 if (!REG_P (target) || HARD_REGISTER_P (target))
4072 valreg = avoid_likely_spilled_reg (valreg);
4074 /* If TARGET is a MEM in the argument area, and we have
4075 saved part of the argument area, then we can't store
4076 directly into TARGET as it may get overwritten when we
4077 restore the argument save area below. Don't work too
4078 hard though and simply force TARGET to a register if it
4079 is a MEM; the optimizer is quite likely to sort it out. */
4080 if (ACCUMULATE_OUTGOING_ARGS && pass && MEM_P (target))
4081 for (i = 0; i < num_actuals; i++)
4082 if (args[i].save_area)
4084 may_overlap = true;
4085 break;
4088 if (may_overlap)
4089 target = copy_to_reg (valreg);
4090 else
4092 /* TARGET and VALREG cannot be equal at this point
4093 because the latter would not have
4094 REG_FUNCTION_VALUE_P true, while the former would if
4095 it were referring to the same register.
4097 If they refer to the same register, this move will be
4098 a no-op, except when function inlining is being
4099 done. */
4100 emit_move_insn (target, valreg);
4102 /* If we are setting a MEM, this code must be executed.
4103 Since it is emitted after the call insn, sibcall
4104 optimization cannot be performed in that case. */
4105 if (MEM_P (target))
4106 sibcall_failure = 1;
4109 else
4110 target = copy_to_reg (avoid_likely_spilled_reg (valreg));
4112 /* If we promoted this return value, make the proper SUBREG.
4113 TARGET might be const0_rtx here, so be careful. */
4114 if (REG_P (target)
4115 && TYPE_MODE (rettype) != BLKmode
4116 && GET_MODE (target) != TYPE_MODE (rettype))
4118 tree type = rettype;
4119 int unsignedp = TYPE_UNSIGNED (type);
4120 int offset = 0;
4121 machine_mode pmode;
4123 /* Ensure we promote as expected, and get the new unsignedness. */
4124 pmode = promote_function_mode (type, TYPE_MODE (type), &unsignedp,
4125 funtype, 1);
4126 gcc_assert (GET_MODE (target) == pmode);
4128 if ((WORDS_BIG_ENDIAN || BYTES_BIG_ENDIAN)
4129 && (GET_MODE_SIZE (GET_MODE (target))
4130 > GET_MODE_SIZE (TYPE_MODE (type))))
4132 offset = GET_MODE_SIZE (GET_MODE (target))
4133 - GET_MODE_SIZE (TYPE_MODE (type));
4134 if (! BYTES_BIG_ENDIAN)
4135 offset = (offset / UNITS_PER_WORD) * UNITS_PER_WORD;
4136 else if (! WORDS_BIG_ENDIAN)
4137 offset %= UNITS_PER_WORD;
4140 target = gen_rtx_SUBREG (TYPE_MODE (type), target, offset);
4141 SUBREG_PROMOTED_VAR_P (target) = 1;
4142 SUBREG_PROMOTED_SET (target, unsignedp);
4145 /* If size of args is variable or this was a constructor call for a stack
4146 argument, restore saved stack-pointer value. */
4148 if (old_stack_level)
4150 rtx_insn *prev = get_last_insn ();
4152 emit_stack_restore (SAVE_BLOCK, old_stack_level);
4153 stack_pointer_delta = old_stack_pointer_delta;
4155 fixup_args_size_notes (prev, get_last_insn (), stack_pointer_delta);
4157 pending_stack_adjust = old_pending_adj;
4158 old_stack_allocated = stack_pointer_delta - pending_stack_adjust;
4159 stack_arg_under_construction = old_stack_arg_under_construction;
4160 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
4161 stack_usage_map = initial_stack_usage_map;
4162 sibcall_failure = 1;
4164 else if (ACCUMULATE_OUTGOING_ARGS && pass)
4166 #ifdef REG_PARM_STACK_SPACE
4167 if (save_area)
4168 restore_fixed_argument_area (save_area, argblock,
4169 high_to_save, low_to_save);
4170 #endif
4172 /* If we saved any argument areas, restore them. */
4173 for (i = 0; i < num_actuals; i++)
4174 if (args[i].save_area)
4176 machine_mode save_mode = GET_MODE (args[i].save_area);
4177 rtx stack_area
4178 = gen_rtx_MEM (save_mode,
4179 memory_address (save_mode,
4180 XEXP (args[i].stack_slot, 0)));
4182 if (save_mode != BLKmode)
4183 emit_move_insn (stack_area, args[i].save_area);
4184 else
4185 emit_block_move (stack_area, args[i].save_area,
4186 GEN_INT (args[i].locate.size.constant),
4187 BLOCK_OP_CALL_PARM);
4190 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
4191 stack_usage_map = initial_stack_usage_map;
4194 /* If this was alloca, record the new stack level. */
4195 if (flags & ECF_MAY_BE_ALLOCA)
4196 record_new_stack_level ();
4198 /* Free up storage we no longer need. */
4199 for (i = 0; i < num_actuals; ++i)
4200 free (args[i].aligned_regs);
4202 targetm.calls.end_call_args ();
4204 insns = get_insns ();
4205 end_sequence ();
4207 if (pass == 0)
4209 tail_call_insns = insns;
4211 /* Restore the pending stack adjustment now that we have
4212 finished generating the sibling call sequence. */
4214 restore_pending_stack_adjust (&save);
4216 /* Prepare arg structure for next iteration. */
4217 for (i = 0; i < num_actuals; i++)
4219 args[i].value = 0;
4220 args[i].aligned_regs = 0;
4221 args[i].stack = 0;
4224 sbitmap_free (stored_args_map);
4225 internal_arg_pointer_exp_state.scan_start = NULL;
4226 internal_arg_pointer_exp_state.cache.release ();
4228 else
4230 normal_call_insns = insns;
4232 /* Verify that we've deallocated all the stack we used. */
4233 gcc_assert ((flags & ECF_NORETURN)
4234 || (old_stack_allocated
4235 == stack_pointer_delta - pending_stack_adjust));
4238 /* If something prevents making this a sibling call,
4239 zero out the sequence. */
4240 if (sibcall_failure)
4241 tail_call_insns = NULL;
4242 else
4243 break;
4246 /* If tail call production succeeded, we need to remove REG_EQUIV notes on
4247 arguments too, as argument area is now clobbered by the call. */
4248 if (tail_call_insns)
4250 emit_insn (tail_call_insns);
4251 crtl->tail_call_emit = true;
4253 else
4255 emit_insn (normal_call_insns);
4256 if (try_tail_call)
4257 /* Ideally we'd emit a message for all of the ways that it could
4258 have failed. */
4259 maybe_complain_about_tail_call (exp, "tail call production failed");
4262 currently_expanding_call--;
4264 free (stack_usage_map_buf);
4265 free (args);
4267 /* Join result with returned bounds so caller may use them if needed. */
4268 target = chkp_join_splitted_slot (target, valbnd);
4270 return target;
4273 /* A sibling call sequence invalidates any REG_EQUIV notes made for
4274 this function's incoming arguments.
4276 At the start of RTL generation we know the only REG_EQUIV notes
4277 in the rtl chain are those for incoming arguments, so we can look
4278 for REG_EQUIV notes between the start of the function and the
4279 NOTE_INSN_FUNCTION_BEG.
4281 This is (slight) overkill. We could keep track of the highest
4282 argument we clobber and be more selective in removing notes, but it
4283 does not seem to be worth the effort. */
4285 void
4286 fixup_tail_calls (void)
4288 rtx_insn *insn;
4290 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
4292 rtx note;
4294 /* There are never REG_EQUIV notes for the incoming arguments
4295 after the NOTE_INSN_FUNCTION_BEG note, so stop if we see it. */
4296 if (NOTE_P (insn)
4297 && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG)
4298 break;
4300 note = find_reg_note (insn, REG_EQUIV, 0);
4301 if (note)
4302 remove_note (insn, note);
4303 note = find_reg_note (insn, REG_EQUIV, 0);
4304 gcc_assert (!note);
4308 /* Traverse a list of TYPES and expand all complex types into their
4309 components. */
4310 static tree
4311 split_complex_types (tree types)
4313 tree p;
4315 /* Before allocating memory, check for the common case of no complex. */
4316 for (p = types; p; p = TREE_CHAIN (p))
4318 tree type = TREE_VALUE (p);
4319 if (TREE_CODE (type) == COMPLEX_TYPE
4320 && targetm.calls.split_complex_arg (type))
4321 goto found;
4323 return types;
4325 found:
4326 types = copy_list (types);
4328 for (p = types; p; p = TREE_CHAIN (p))
4330 tree complex_type = TREE_VALUE (p);
4332 if (TREE_CODE (complex_type) == COMPLEX_TYPE
4333 && targetm.calls.split_complex_arg (complex_type))
4335 tree next, imag;
4337 /* Rewrite complex type with component type. */
4338 TREE_VALUE (p) = TREE_TYPE (complex_type);
4339 next = TREE_CHAIN (p);
4341 /* Add another component type for the imaginary part. */
4342 imag = build_tree_list (NULL_TREE, TREE_VALUE (p));
4343 TREE_CHAIN (p) = imag;
4344 TREE_CHAIN (imag) = next;
4346 /* Skip the newly created node. */
4347 p = TREE_CHAIN (p);
4351 return types;
4354 /* Output a library call to function FUN (a SYMBOL_REF rtx).
4355 The RETVAL parameter specifies whether return value needs to be saved, other
4356 parameters are documented in the emit_library_call function below. */
4358 static rtx
4359 emit_library_call_value_1 (int retval, rtx orgfun, rtx value,
4360 enum libcall_type fn_type,
4361 machine_mode outmode, int nargs, va_list p)
4363 /* Total size in bytes of all the stack-parms scanned so far. */
4364 struct args_size args_size;
4365 /* Size of arguments before any adjustments (such as rounding). */
4366 struct args_size original_args_size;
4367 int argnum;
4368 rtx fun;
4369 /* Todo, choose the correct decl type of orgfun. Sadly this information
4370 isn't present here, so we default to native calling abi here. */
4371 tree fndecl ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */
4372 tree fntype ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */
4373 int count;
4374 rtx argblock = 0;
4375 CUMULATIVE_ARGS args_so_far_v;
4376 cumulative_args_t args_so_far;
4377 struct arg
4379 rtx value;
4380 machine_mode mode;
4381 rtx reg;
4382 int partial;
4383 struct locate_and_pad_arg_data locate;
4384 rtx save_area;
4386 struct arg *argvec;
4387 int old_inhibit_defer_pop = inhibit_defer_pop;
4388 rtx call_fusage = 0;
4389 rtx mem_value = 0;
4390 rtx valreg;
4391 int pcc_struct_value = 0;
4392 int struct_value_size = 0;
4393 int flags;
4394 int reg_parm_stack_space = 0;
4395 int needed;
4396 rtx_insn *before_call;
4397 bool have_push_fusage;
4398 tree tfom; /* type_for_mode (outmode, 0) */
4400 #ifdef REG_PARM_STACK_SPACE
4401 /* Define the boundary of the register parm stack space that needs to be
4402 save, if any. */
4403 int low_to_save = 0, high_to_save = 0;
4404 rtx save_area = 0; /* Place that it is saved. */
4405 #endif
4407 /* Size of the stack reserved for parameter registers. */
4408 int initial_highest_arg_in_use = highest_outgoing_arg_in_use;
4409 char *initial_stack_usage_map = stack_usage_map;
4410 char *stack_usage_map_buf = NULL;
4412 rtx struct_value = targetm.calls.struct_value_rtx (0, 0);
4414 #ifdef REG_PARM_STACK_SPACE
4415 reg_parm_stack_space = REG_PARM_STACK_SPACE ((tree) 0);
4416 #endif
4418 /* By default, library functions cannot throw. */
4419 flags = ECF_NOTHROW;
4421 switch (fn_type)
4423 case LCT_NORMAL:
4424 break;
4425 case LCT_CONST:
4426 flags |= ECF_CONST;
4427 break;
4428 case LCT_PURE:
4429 flags |= ECF_PURE;
4430 break;
4431 case LCT_NORETURN:
4432 flags |= ECF_NORETURN;
4433 break;
4434 case LCT_THROW:
4435 flags &= ~ECF_NOTHROW;
4436 break;
4437 case LCT_RETURNS_TWICE:
4438 flags = ECF_RETURNS_TWICE;
4439 break;
4441 fun = orgfun;
4443 /* Ensure current function's preferred stack boundary is at least
4444 what we need. */
4445 if (crtl->preferred_stack_boundary < PREFERRED_STACK_BOUNDARY)
4446 crtl->preferred_stack_boundary = PREFERRED_STACK_BOUNDARY;
4448 /* If this kind of value comes back in memory,
4449 decide where in memory it should come back. */
4450 if (outmode != VOIDmode)
4452 tfom = lang_hooks.types.type_for_mode (outmode, 0);
4453 if (aggregate_value_p (tfom, 0))
4455 #ifdef PCC_STATIC_STRUCT_RETURN
4456 rtx pointer_reg
4457 = hard_function_value (build_pointer_type (tfom), 0, 0, 0);
4458 mem_value = gen_rtx_MEM (outmode, pointer_reg);
4459 pcc_struct_value = 1;
4460 if (value == 0)
4461 value = gen_reg_rtx (outmode);
4462 #else /* not PCC_STATIC_STRUCT_RETURN */
4463 struct_value_size = GET_MODE_SIZE (outmode);
4464 if (value != 0 && MEM_P (value))
4465 mem_value = value;
4466 else
4467 mem_value = assign_temp (tfom, 1, 1);
4468 #endif
4469 /* This call returns a big structure. */
4470 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
4473 else
4474 tfom = void_type_node;
4476 /* ??? Unfinished: must pass the memory address as an argument. */
4478 /* Copy all the libcall-arguments out of the varargs data
4479 and into a vector ARGVEC.
4481 Compute how to pass each argument. We only support a very small subset
4482 of the full argument passing conventions to limit complexity here since
4483 library functions shouldn't have many args. */
4485 argvec = XALLOCAVEC (struct arg, nargs + 1);
4486 memset (argvec, 0, (nargs + 1) * sizeof (struct arg));
4488 #ifdef INIT_CUMULATIVE_LIBCALL_ARGS
4489 INIT_CUMULATIVE_LIBCALL_ARGS (args_so_far_v, outmode, fun);
4490 #else
4491 INIT_CUMULATIVE_ARGS (args_so_far_v, NULL_TREE, fun, 0, nargs);
4492 #endif
4493 args_so_far = pack_cumulative_args (&args_so_far_v);
4495 args_size.constant = 0;
4496 args_size.var = 0;
4498 count = 0;
4500 push_temp_slots ();
4502 /* If there's a structure value address to be passed,
4503 either pass it in the special place, or pass it as an extra argument. */
4504 if (mem_value && struct_value == 0 && ! pcc_struct_value)
4506 rtx addr = XEXP (mem_value, 0);
4508 nargs++;
4510 /* Make sure it is a reasonable operand for a move or push insn. */
4511 if (!REG_P (addr) && !MEM_P (addr)
4512 && !(CONSTANT_P (addr)
4513 && targetm.legitimate_constant_p (Pmode, addr)))
4514 addr = force_operand (addr, NULL_RTX);
4516 argvec[count].value = addr;
4517 argvec[count].mode = Pmode;
4518 argvec[count].partial = 0;
4520 argvec[count].reg = targetm.calls.function_arg (args_so_far,
4521 Pmode, NULL_TREE, true);
4522 gcc_assert (targetm.calls.arg_partial_bytes (args_so_far, Pmode,
4523 NULL_TREE, 1) == 0);
4525 locate_and_pad_parm (Pmode, NULL_TREE,
4526 #ifdef STACK_PARMS_IN_REG_PARM_AREA
4528 #else
4529 argvec[count].reg != 0,
4530 #endif
4531 reg_parm_stack_space, 0,
4532 NULL_TREE, &args_size, &argvec[count].locate);
4534 if (argvec[count].reg == 0 || argvec[count].partial != 0
4535 || reg_parm_stack_space > 0)
4536 args_size.constant += argvec[count].locate.size.constant;
4538 targetm.calls.function_arg_advance (args_so_far, Pmode, (tree) 0, true);
4540 count++;
4543 for (; count < nargs; count++)
4545 rtx val = va_arg (p, rtx);
4546 machine_mode mode = (machine_mode) va_arg (p, int);
4547 int unsigned_p = 0;
4549 /* We cannot convert the arg value to the mode the library wants here;
4550 must do it earlier where we know the signedness of the arg. */
4551 gcc_assert (mode != BLKmode
4552 && (GET_MODE (val) == mode || GET_MODE (val) == VOIDmode));
4554 /* Make sure it is a reasonable operand for a move or push insn. */
4555 if (!REG_P (val) && !MEM_P (val)
4556 && !(CONSTANT_P (val) && targetm.legitimate_constant_p (mode, val)))
4557 val = force_operand (val, NULL_RTX);
4559 if (pass_by_reference (&args_so_far_v, mode, NULL_TREE, 1))
4561 rtx slot;
4562 int must_copy
4563 = !reference_callee_copied (&args_so_far_v, mode, NULL_TREE, 1);
4565 /* If this was a CONST function, it is now PURE since it now
4566 reads memory. */
4567 if (flags & ECF_CONST)
4569 flags &= ~ECF_CONST;
4570 flags |= ECF_PURE;
4573 if (MEM_P (val) && !must_copy)
4575 tree val_expr = MEM_EXPR (val);
4576 if (val_expr)
4577 mark_addressable (val_expr);
4578 slot = val;
4580 else
4582 slot = assign_temp (lang_hooks.types.type_for_mode (mode, 0),
4583 1, 1);
4584 emit_move_insn (slot, val);
4587 call_fusage = gen_rtx_EXPR_LIST (VOIDmode,
4588 gen_rtx_USE (VOIDmode, slot),
4589 call_fusage);
4590 if (must_copy)
4591 call_fusage = gen_rtx_EXPR_LIST (VOIDmode,
4592 gen_rtx_CLOBBER (VOIDmode,
4593 slot),
4594 call_fusage);
4596 mode = Pmode;
4597 val = force_operand (XEXP (slot, 0), NULL_RTX);
4600 mode = promote_function_mode (NULL_TREE, mode, &unsigned_p, NULL_TREE, 0);
4601 argvec[count].mode = mode;
4602 argvec[count].value = convert_modes (mode, GET_MODE (val), val, unsigned_p);
4603 argvec[count].reg = targetm.calls.function_arg (args_so_far, mode,
4604 NULL_TREE, true);
4606 argvec[count].partial
4607 = targetm.calls.arg_partial_bytes (args_so_far, mode, NULL_TREE, 1);
4609 if (argvec[count].reg == 0
4610 || argvec[count].partial != 0
4611 || reg_parm_stack_space > 0)
4613 locate_and_pad_parm (mode, NULL_TREE,
4614 #ifdef STACK_PARMS_IN_REG_PARM_AREA
4616 #else
4617 argvec[count].reg != 0,
4618 #endif
4619 reg_parm_stack_space, argvec[count].partial,
4620 NULL_TREE, &args_size, &argvec[count].locate);
4621 args_size.constant += argvec[count].locate.size.constant;
4622 gcc_assert (!argvec[count].locate.size.var);
4624 #ifdef BLOCK_REG_PADDING
4625 else
4626 /* The argument is passed entirely in registers. See at which
4627 end it should be padded. */
4628 argvec[count].locate.where_pad =
4629 BLOCK_REG_PADDING (mode, NULL_TREE,
4630 GET_MODE_SIZE (mode) <= UNITS_PER_WORD);
4631 #endif
4633 targetm.calls.function_arg_advance (args_so_far, mode, (tree) 0, true);
4636 /* If this machine requires an external definition for library
4637 functions, write one out. */
4638 assemble_external_libcall (fun);
4640 original_args_size = args_size;
4641 args_size.constant = (((args_size.constant
4642 + stack_pointer_delta
4643 + STACK_BYTES - 1)
4644 / STACK_BYTES
4645 * STACK_BYTES)
4646 - stack_pointer_delta);
4648 args_size.constant = MAX (args_size.constant,
4649 reg_parm_stack_space);
4651 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
4652 args_size.constant -= reg_parm_stack_space;
4654 if (args_size.constant > crtl->outgoing_args_size)
4655 crtl->outgoing_args_size = args_size.constant;
4657 if (flag_stack_usage_info && !ACCUMULATE_OUTGOING_ARGS)
4659 int pushed = args_size.constant + pending_stack_adjust;
4660 if (pushed > current_function_pushed_stack_size)
4661 current_function_pushed_stack_size = pushed;
4664 if (ACCUMULATE_OUTGOING_ARGS)
4666 /* Since the stack pointer will never be pushed, it is possible for
4667 the evaluation of a parm to clobber something we have already
4668 written to the stack. Since most function calls on RISC machines
4669 do not use the stack, this is uncommon, but must work correctly.
4671 Therefore, we save any area of the stack that was already written
4672 and that we are using. Here we set up to do this by making a new
4673 stack usage map from the old one.
4675 Another approach might be to try to reorder the argument
4676 evaluations to avoid this conflicting stack usage. */
4678 needed = args_size.constant;
4680 /* Since we will be writing into the entire argument area, the
4681 map must be allocated for its entire size, not just the part that
4682 is the responsibility of the caller. */
4683 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
4684 needed += reg_parm_stack_space;
4686 if (ARGS_GROW_DOWNWARD)
4687 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use,
4688 needed + 1);
4689 else
4690 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use, needed);
4692 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use);
4693 stack_usage_map = stack_usage_map_buf;
4695 if (initial_highest_arg_in_use)
4696 memcpy (stack_usage_map, initial_stack_usage_map,
4697 initial_highest_arg_in_use);
4699 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use)
4700 memset (&stack_usage_map[initial_highest_arg_in_use], 0,
4701 highest_outgoing_arg_in_use - initial_highest_arg_in_use);
4702 needed = 0;
4704 /* We must be careful to use virtual regs before they're instantiated,
4705 and real regs afterwards. Loop optimization, for example, can create
4706 new libcalls after we've instantiated the virtual regs, and if we
4707 use virtuals anyway, they won't match the rtl patterns. */
4709 if (virtuals_instantiated)
4710 argblock = plus_constant (Pmode, stack_pointer_rtx,
4711 STACK_POINTER_OFFSET);
4712 else
4713 argblock = virtual_outgoing_args_rtx;
4715 else
4717 if (!PUSH_ARGS)
4718 argblock = push_block (GEN_INT (args_size.constant), 0, 0);
4721 /* We push args individually in reverse order, perform stack alignment
4722 before the first push (the last arg). */
4723 if (argblock == 0)
4724 anti_adjust_stack (GEN_INT (args_size.constant
4725 - original_args_size.constant));
4727 argnum = nargs - 1;
4729 #ifdef REG_PARM_STACK_SPACE
4730 if (ACCUMULATE_OUTGOING_ARGS)
4732 /* The argument list is the property of the called routine and it
4733 may clobber it. If the fixed area has been used for previous
4734 parameters, we must save and restore it. */
4735 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock,
4736 &low_to_save, &high_to_save);
4738 #endif
4740 /* When expanding a normal call, args are stored in push order,
4741 which is the reverse of what we have here. */
4742 bool any_regs = false;
4743 for (int i = nargs; i-- > 0; )
4744 if (argvec[i].reg != NULL_RTX)
4746 targetm.calls.call_args (argvec[i].reg, NULL_TREE);
4747 any_regs = true;
4749 if (!any_regs)
4750 targetm.calls.call_args (pc_rtx, NULL_TREE);
4752 /* Push the args that need to be pushed. */
4754 have_push_fusage = false;
4756 /* ARGNUM indexes the ARGVEC array in the order in which the arguments
4757 are to be pushed. */
4758 for (count = 0; count < nargs; count++, argnum--)
4760 machine_mode mode = argvec[argnum].mode;
4761 rtx val = argvec[argnum].value;
4762 rtx reg = argvec[argnum].reg;
4763 int partial = argvec[argnum].partial;
4764 unsigned int parm_align = argvec[argnum].locate.boundary;
4765 int lower_bound = 0, upper_bound = 0, i;
4767 if (! (reg != 0 && partial == 0))
4769 rtx use;
4771 if (ACCUMULATE_OUTGOING_ARGS)
4773 /* If this is being stored into a pre-allocated, fixed-size,
4774 stack area, save any previous data at that location. */
4776 if (ARGS_GROW_DOWNWARD)
4778 /* stack_slot is negative, but we want to index stack_usage_map
4779 with positive values. */
4780 upper_bound = -argvec[argnum].locate.slot_offset.constant + 1;
4781 lower_bound = upper_bound - argvec[argnum].locate.size.constant;
4783 else
4785 lower_bound = argvec[argnum].locate.slot_offset.constant;
4786 upper_bound = lower_bound + argvec[argnum].locate.size.constant;
4789 i = lower_bound;
4790 /* Don't worry about things in the fixed argument area;
4791 it has already been saved. */
4792 if (i < reg_parm_stack_space)
4793 i = reg_parm_stack_space;
4794 while (i < upper_bound && stack_usage_map[i] == 0)
4795 i++;
4797 if (i < upper_bound)
4799 /* We need to make a save area. */
4800 unsigned int size
4801 = argvec[argnum].locate.size.constant * BITS_PER_UNIT;
4802 machine_mode save_mode
4803 = mode_for_size (size, MODE_INT, 1);
4804 rtx adr
4805 = plus_constant (Pmode, argblock,
4806 argvec[argnum].locate.offset.constant);
4807 rtx stack_area
4808 = gen_rtx_MEM (save_mode, memory_address (save_mode, adr));
4810 if (save_mode == BLKmode)
4812 argvec[argnum].save_area
4813 = assign_stack_temp (BLKmode,
4814 argvec[argnum].locate.size.constant
4817 emit_block_move (validize_mem
4818 (copy_rtx (argvec[argnum].save_area)),
4819 stack_area,
4820 GEN_INT (argvec[argnum].locate.size.constant),
4821 BLOCK_OP_CALL_PARM);
4823 else
4825 argvec[argnum].save_area = gen_reg_rtx (save_mode);
4827 emit_move_insn (argvec[argnum].save_area, stack_area);
4832 emit_push_insn (val, mode, NULL_TREE, NULL_RTX, parm_align,
4833 partial, reg, 0, argblock,
4834 GEN_INT (argvec[argnum].locate.offset.constant),
4835 reg_parm_stack_space,
4836 ARGS_SIZE_RTX (argvec[argnum].locate.alignment_pad), false);
4838 /* Now mark the segment we just used. */
4839 if (ACCUMULATE_OUTGOING_ARGS)
4840 for (i = lower_bound; i < upper_bound; i++)
4841 stack_usage_map[i] = 1;
4843 NO_DEFER_POP;
4845 /* Indicate argument access so that alias.c knows that these
4846 values are live. */
4847 if (argblock)
4848 use = plus_constant (Pmode, argblock,
4849 argvec[argnum].locate.offset.constant);
4850 else if (have_push_fusage)
4851 continue;
4852 else
4854 /* When arguments are pushed, trying to tell alias.c where
4855 exactly this argument is won't work, because the
4856 auto-increment causes confusion. So we merely indicate
4857 that we access something with a known mode somewhere on
4858 the stack. */
4859 use = gen_rtx_PLUS (Pmode, stack_pointer_rtx,
4860 gen_rtx_SCRATCH (Pmode));
4861 have_push_fusage = true;
4863 use = gen_rtx_MEM (argvec[argnum].mode, use);
4864 use = gen_rtx_USE (VOIDmode, use);
4865 call_fusage = gen_rtx_EXPR_LIST (VOIDmode, use, call_fusage);
4869 argnum = nargs - 1;
4871 fun = prepare_call_address (NULL, fun, NULL, &call_fusage, 0, 0);
4873 /* Now load any reg parms into their regs. */
4875 /* ARGNUM indexes the ARGVEC array in the order in which the arguments
4876 are to be pushed. */
4877 for (count = 0; count < nargs; count++, argnum--)
4879 machine_mode mode = argvec[argnum].mode;
4880 rtx val = argvec[argnum].value;
4881 rtx reg = argvec[argnum].reg;
4882 int partial = argvec[argnum].partial;
4883 #ifdef BLOCK_REG_PADDING
4884 int size = 0;
4885 #endif
4887 /* Handle calls that pass values in multiple non-contiguous
4888 locations. The PA64 has examples of this for library calls. */
4889 if (reg != 0 && GET_CODE (reg) == PARALLEL)
4890 emit_group_load (reg, val, NULL_TREE, GET_MODE_SIZE (mode));
4891 else if (reg != 0 && partial == 0)
4893 emit_move_insn (reg, val);
4894 #ifdef BLOCK_REG_PADDING
4895 size = GET_MODE_SIZE (argvec[argnum].mode);
4897 /* Copied from load_register_parameters. */
4899 /* Handle case where we have a value that needs shifting
4900 up to the msb. eg. a QImode value and we're padding
4901 upward on a BYTES_BIG_ENDIAN machine. */
4902 if (size < UNITS_PER_WORD
4903 && (argvec[argnum].locate.where_pad
4904 == (BYTES_BIG_ENDIAN ? upward : downward)))
4906 rtx x;
4907 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
4909 /* Assigning REG here rather than a temp makes CALL_FUSAGE
4910 report the whole reg as used. Strictly speaking, the
4911 call only uses SIZE bytes at the msb end, but it doesn't
4912 seem worth generating rtl to say that. */
4913 reg = gen_rtx_REG (word_mode, REGNO (reg));
4914 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1);
4915 if (x != reg)
4916 emit_move_insn (reg, x);
4918 #endif
4921 NO_DEFER_POP;
4924 /* Any regs containing parms remain in use through the call. */
4925 for (count = 0; count < nargs; count++)
4927 rtx reg = argvec[count].reg;
4928 if (reg != 0 && GET_CODE (reg) == PARALLEL)
4929 use_group_regs (&call_fusage, reg);
4930 else if (reg != 0)
4932 int partial = argvec[count].partial;
4933 if (partial)
4935 int nregs;
4936 gcc_assert (partial % UNITS_PER_WORD == 0);
4937 nregs = partial / UNITS_PER_WORD;
4938 use_regs (&call_fusage, REGNO (reg), nregs);
4940 else
4941 use_reg (&call_fusage, reg);
4945 /* Pass the function the address in which to return a structure value. */
4946 if (mem_value != 0 && struct_value != 0 && ! pcc_struct_value)
4948 emit_move_insn (struct_value,
4949 force_reg (Pmode,
4950 force_operand (XEXP (mem_value, 0),
4951 NULL_RTX)));
4952 if (REG_P (struct_value))
4953 use_reg (&call_fusage, struct_value);
4956 /* Don't allow popping to be deferred, since then
4957 cse'ing of library calls could delete a call and leave the pop. */
4958 NO_DEFER_POP;
4959 valreg = (mem_value == 0 && outmode != VOIDmode
4960 ? hard_libcall_value (outmode, orgfun) : NULL_RTX);
4962 /* Stack must be properly aligned now. */
4963 gcc_assert (!(stack_pointer_delta
4964 & (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT - 1)));
4966 before_call = get_last_insn ();
4968 /* We pass the old value of inhibit_defer_pop + 1 to emit_call_1, which
4969 will set inhibit_defer_pop to that value. */
4970 /* The return type is needed to decide how many bytes the function pops.
4971 Signedness plays no role in that, so for simplicity, we pretend it's
4972 always signed. We also assume that the list of arguments passed has
4973 no impact, so we pretend it is unknown. */
4975 emit_call_1 (fun, NULL,
4976 get_identifier (XSTR (orgfun, 0)),
4977 build_function_type (tfom, NULL_TREE),
4978 original_args_size.constant, args_size.constant,
4979 struct_value_size,
4980 targetm.calls.function_arg (args_so_far,
4981 VOIDmode, void_type_node, true),
4982 valreg,
4983 old_inhibit_defer_pop + 1, call_fusage, flags, args_so_far);
4985 if (flag_ipa_ra)
4987 rtx datum = orgfun;
4988 gcc_assert (GET_CODE (datum) == SYMBOL_REF);
4989 rtx_call_insn *last = last_call_insn ();
4990 add_reg_note (last, REG_CALL_DECL, datum);
4993 /* Right-shift returned value if necessary. */
4994 if (!pcc_struct_value
4995 && TYPE_MODE (tfom) != BLKmode
4996 && targetm.calls.return_in_msb (tfom))
4998 shift_return_value (TYPE_MODE (tfom), false, valreg);
4999 valreg = gen_rtx_REG (TYPE_MODE (tfom), REGNO (valreg));
5002 targetm.calls.end_call_args ();
5004 /* For calls to `setjmp', etc., inform function.c:setjmp_warnings
5005 that it should complain if nonvolatile values are live. For
5006 functions that cannot return, inform flow that control does not
5007 fall through. */
5008 if (flags & ECF_NORETURN)
5010 /* The barrier note must be emitted
5011 immediately after the CALL_INSN. Some ports emit more than
5012 just a CALL_INSN above, so we must search for it here. */
5013 rtx_insn *last = get_last_insn ();
5014 while (!CALL_P (last))
5016 last = PREV_INSN (last);
5017 /* There was no CALL_INSN? */
5018 gcc_assert (last != before_call);
5021 emit_barrier_after (last);
5024 /* Consider that "regular" libcalls, i.e. all of them except for LCT_THROW
5025 and LCT_RETURNS_TWICE, cannot perform non-local gotos. */
5026 if (flags & ECF_NOTHROW)
5028 rtx_insn *last = get_last_insn ();
5029 while (!CALL_P (last))
5031 last = PREV_INSN (last);
5032 /* There was no CALL_INSN? */
5033 gcc_assert (last != before_call);
5036 make_reg_eh_region_note_nothrow_nononlocal (last);
5039 /* Now restore inhibit_defer_pop to its actual original value. */
5040 OK_DEFER_POP;
5042 pop_temp_slots ();
5044 /* Copy the value to the right place. */
5045 if (outmode != VOIDmode && retval)
5047 if (mem_value)
5049 if (value == 0)
5050 value = mem_value;
5051 if (value != mem_value)
5052 emit_move_insn (value, mem_value);
5054 else if (GET_CODE (valreg) == PARALLEL)
5056 if (value == 0)
5057 value = gen_reg_rtx (outmode);
5058 emit_group_store (value, valreg, NULL_TREE, GET_MODE_SIZE (outmode));
5060 else
5062 /* Convert to the proper mode if a promotion has been active. */
5063 if (GET_MODE (valreg) != outmode)
5065 int unsignedp = TYPE_UNSIGNED (tfom);
5067 gcc_assert (promote_function_mode (tfom, outmode, &unsignedp,
5068 fndecl ? TREE_TYPE (fndecl) : fntype, 1)
5069 == GET_MODE (valreg));
5070 valreg = convert_modes (outmode, GET_MODE (valreg), valreg, 0);
5073 if (value != 0)
5074 emit_move_insn (value, valreg);
5075 else
5076 value = valreg;
5080 if (ACCUMULATE_OUTGOING_ARGS)
5082 #ifdef REG_PARM_STACK_SPACE
5083 if (save_area)
5084 restore_fixed_argument_area (save_area, argblock,
5085 high_to_save, low_to_save);
5086 #endif
5088 /* If we saved any argument areas, restore them. */
5089 for (count = 0; count < nargs; count++)
5090 if (argvec[count].save_area)
5092 machine_mode save_mode = GET_MODE (argvec[count].save_area);
5093 rtx adr = plus_constant (Pmode, argblock,
5094 argvec[count].locate.offset.constant);
5095 rtx stack_area = gen_rtx_MEM (save_mode,
5096 memory_address (save_mode, adr));
5098 if (save_mode == BLKmode)
5099 emit_block_move (stack_area,
5100 validize_mem
5101 (copy_rtx (argvec[count].save_area)),
5102 GEN_INT (argvec[count].locate.size.constant),
5103 BLOCK_OP_CALL_PARM);
5104 else
5105 emit_move_insn (stack_area, argvec[count].save_area);
5108 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
5109 stack_usage_map = initial_stack_usage_map;
5112 free (stack_usage_map_buf);
5114 return value;
5118 /* Output a library call to function FUN (a SYMBOL_REF rtx)
5119 (emitting the queue unless NO_QUEUE is nonzero),
5120 for a value of mode OUTMODE,
5121 with NARGS different arguments, passed as alternating rtx values
5122 and machine_modes to convert them to.
5124 FN_TYPE should be LCT_NORMAL for `normal' calls, LCT_CONST for
5125 `const' calls, LCT_PURE for `pure' calls, or other LCT_ value for
5126 other types of library calls. */
5128 void
5129 emit_library_call (rtx orgfun, enum libcall_type fn_type,
5130 machine_mode outmode, int nargs, ...)
5132 va_list p;
5134 va_start (p, nargs);
5135 emit_library_call_value_1 (0, orgfun, NULL_RTX, fn_type, outmode, nargs, p);
5136 va_end (p);
5139 /* Like emit_library_call except that an extra argument, VALUE,
5140 comes second and says where to store the result.
5141 (If VALUE is zero, this function chooses a convenient way
5142 to return the value.
5144 This function returns an rtx for where the value is to be found.
5145 If VALUE is nonzero, VALUE is returned. */
5148 emit_library_call_value (rtx orgfun, rtx value,
5149 enum libcall_type fn_type,
5150 machine_mode outmode, int nargs, ...)
5152 rtx result;
5153 va_list p;
5155 va_start (p, nargs);
5156 result = emit_library_call_value_1 (1, orgfun, value, fn_type, outmode,
5157 nargs, p);
5158 va_end (p);
5160 return result;
5164 /* Store pointer bounds argument ARG into Bounds Table entry
5165 associated with PARM. */
5166 static void
5167 store_bounds (struct arg_data *arg, struct arg_data *parm)
5169 rtx slot = NULL, ptr = NULL, addr = NULL;
5171 /* We may pass bounds not associated with any pointer. */
5172 if (!parm)
5174 gcc_assert (arg->special_slot);
5175 slot = arg->special_slot;
5176 ptr = const0_rtx;
5178 /* Find pointer associated with bounds and where it is
5179 passed. */
5180 else
5182 if (!parm->reg)
5184 gcc_assert (!arg->special_slot);
5186 addr = adjust_address (parm->stack, Pmode, arg->pointer_offset);
5188 else if (REG_P (parm->reg))
5190 gcc_assert (arg->special_slot);
5191 slot = arg->special_slot;
5193 if (MEM_P (parm->value))
5194 addr = adjust_address (parm->value, Pmode, arg->pointer_offset);
5195 else if (REG_P (parm->value))
5196 ptr = gen_rtx_SUBREG (Pmode, parm->value, arg->pointer_offset);
5197 else
5199 gcc_assert (!arg->pointer_offset);
5200 ptr = parm->value;
5203 else
5205 gcc_assert (GET_CODE (parm->reg) == PARALLEL);
5207 gcc_assert (arg->special_slot);
5208 slot = arg->special_slot;
5210 if (parm->parallel_value)
5211 ptr = chkp_get_value_with_offs (parm->parallel_value,
5212 GEN_INT (arg->pointer_offset));
5213 else
5214 gcc_unreachable ();
5218 /* Expand bounds. */
5219 if (!arg->value)
5220 arg->value = expand_normal (arg->tree_value);
5222 targetm.calls.store_bounds_for_arg (ptr, addr, arg->value, slot);
5225 /* Store a single argument for a function call
5226 into the register or memory area where it must be passed.
5227 *ARG describes the argument value and where to pass it.
5229 ARGBLOCK is the address of the stack-block for all the arguments,
5230 or 0 on a machine where arguments are pushed individually.
5232 MAY_BE_ALLOCA nonzero says this could be a call to `alloca'
5233 so must be careful about how the stack is used.
5235 VARIABLE_SIZE nonzero says that this was a variable-sized outgoing
5236 argument stack. This is used if ACCUMULATE_OUTGOING_ARGS to indicate
5237 that we need not worry about saving and restoring the stack.
5239 FNDECL is the declaration of the function we are calling.
5241 Return nonzero if this arg should cause sibcall failure,
5242 zero otherwise. */
5244 static int
5245 store_one_arg (struct arg_data *arg, rtx argblock, int flags,
5246 int variable_size ATTRIBUTE_UNUSED, int reg_parm_stack_space)
5248 tree pval = arg->tree_value;
5249 rtx reg = 0;
5250 int partial = 0;
5251 int used = 0;
5252 int i, lower_bound = 0, upper_bound = 0;
5253 int sibcall_failure = 0;
5255 if (TREE_CODE (pval) == ERROR_MARK)
5256 return 1;
5258 /* Push a new temporary level for any temporaries we make for
5259 this argument. */
5260 push_temp_slots ();
5262 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL))
5264 /* If this is being stored into a pre-allocated, fixed-size, stack area,
5265 save any previous data at that location. */
5266 if (argblock && ! variable_size && arg->stack)
5268 if (ARGS_GROW_DOWNWARD)
5270 /* stack_slot is negative, but we want to index stack_usage_map
5271 with positive values. */
5272 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS)
5273 upper_bound = -INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1)) + 1;
5274 else
5275 upper_bound = 0;
5277 lower_bound = upper_bound - arg->locate.size.constant;
5279 else
5281 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS)
5282 lower_bound = INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1));
5283 else
5284 lower_bound = 0;
5286 upper_bound = lower_bound + arg->locate.size.constant;
5289 i = lower_bound;
5290 /* Don't worry about things in the fixed argument area;
5291 it has already been saved. */
5292 if (i < reg_parm_stack_space)
5293 i = reg_parm_stack_space;
5294 while (i < upper_bound && stack_usage_map[i] == 0)
5295 i++;
5297 if (i < upper_bound)
5299 /* We need to make a save area. */
5300 unsigned int size = arg->locate.size.constant * BITS_PER_UNIT;
5301 machine_mode save_mode = mode_for_size (size, MODE_INT, 1);
5302 rtx adr = memory_address (save_mode, XEXP (arg->stack_slot, 0));
5303 rtx stack_area = gen_rtx_MEM (save_mode, adr);
5305 if (save_mode == BLKmode)
5307 arg->save_area
5308 = assign_temp (TREE_TYPE (arg->tree_value), 1, 1);
5309 preserve_temp_slots (arg->save_area);
5310 emit_block_move (validize_mem (copy_rtx (arg->save_area)),
5311 stack_area,
5312 GEN_INT (arg->locate.size.constant),
5313 BLOCK_OP_CALL_PARM);
5315 else
5317 arg->save_area = gen_reg_rtx (save_mode);
5318 emit_move_insn (arg->save_area, stack_area);
5324 /* If this isn't going to be placed on both the stack and in registers,
5325 set up the register and number of words. */
5326 if (! arg->pass_on_stack)
5328 if (flags & ECF_SIBCALL)
5329 reg = arg->tail_call_reg;
5330 else
5331 reg = arg->reg;
5332 partial = arg->partial;
5335 /* Being passed entirely in a register. We shouldn't be called in
5336 this case. */
5337 gcc_assert (reg == 0 || partial != 0);
5339 /* If this arg needs special alignment, don't load the registers
5340 here. */
5341 if (arg->n_aligned_regs != 0)
5342 reg = 0;
5344 /* If this is being passed partially in a register, we can't evaluate
5345 it directly into its stack slot. Otherwise, we can. */
5346 if (arg->value == 0)
5348 /* stack_arg_under_construction is nonzero if a function argument is
5349 being evaluated directly into the outgoing argument list and
5350 expand_call must take special action to preserve the argument list
5351 if it is called recursively.
5353 For scalar function arguments stack_usage_map is sufficient to
5354 determine which stack slots must be saved and restored. Scalar
5355 arguments in general have pass_on_stack == 0.
5357 If this argument is initialized by a function which takes the
5358 address of the argument (a C++ constructor or a C function
5359 returning a BLKmode structure), then stack_usage_map is
5360 insufficient and expand_call must push the stack around the
5361 function call. Such arguments have pass_on_stack == 1.
5363 Note that it is always safe to set stack_arg_under_construction,
5364 but this generates suboptimal code if set when not needed. */
5366 if (arg->pass_on_stack)
5367 stack_arg_under_construction++;
5369 arg->value = expand_expr (pval,
5370 (partial
5371 || TYPE_MODE (TREE_TYPE (pval)) != arg->mode)
5372 ? NULL_RTX : arg->stack,
5373 VOIDmode, EXPAND_STACK_PARM);
5375 /* If we are promoting object (or for any other reason) the mode
5376 doesn't agree, convert the mode. */
5378 if (arg->mode != TYPE_MODE (TREE_TYPE (pval)))
5379 arg->value = convert_modes (arg->mode, TYPE_MODE (TREE_TYPE (pval)),
5380 arg->value, arg->unsignedp);
5382 if (arg->pass_on_stack)
5383 stack_arg_under_construction--;
5386 /* Check for overlap with already clobbered argument area. */
5387 if ((flags & ECF_SIBCALL)
5388 && MEM_P (arg->value)
5389 && mem_overlaps_already_clobbered_arg_p (XEXP (arg->value, 0),
5390 arg->locate.size.constant))
5391 sibcall_failure = 1;
5393 /* Don't allow anything left on stack from computation
5394 of argument to alloca. */
5395 if (flags & ECF_MAY_BE_ALLOCA)
5396 do_pending_stack_adjust ();
5398 if (arg->value == arg->stack)
5399 /* If the value is already in the stack slot, we are done. */
5401 else if (arg->mode != BLKmode)
5403 int size;
5404 unsigned int parm_align;
5406 /* Argument is a scalar, not entirely passed in registers.
5407 (If part is passed in registers, arg->partial says how much
5408 and emit_push_insn will take care of putting it there.)
5410 Push it, and if its size is less than the
5411 amount of space allocated to it,
5412 also bump stack pointer by the additional space.
5413 Note that in C the default argument promotions
5414 will prevent such mismatches. */
5416 size = GET_MODE_SIZE (arg->mode);
5417 /* Compute how much space the push instruction will push.
5418 On many machines, pushing a byte will advance the stack
5419 pointer by a halfword. */
5420 #ifdef PUSH_ROUNDING
5421 size = PUSH_ROUNDING (size);
5422 #endif
5423 used = size;
5425 /* Compute how much space the argument should get:
5426 round up to a multiple of the alignment for arguments. */
5427 if (none != FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)))
5428 used = (((size + PARM_BOUNDARY / BITS_PER_UNIT - 1)
5429 / (PARM_BOUNDARY / BITS_PER_UNIT))
5430 * (PARM_BOUNDARY / BITS_PER_UNIT));
5432 /* Compute the alignment of the pushed argument. */
5433 parm_align = arg->locate.boundary;
5434 if (FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)) == downward)
5436 int pad = used - size;
5437 if (pad)
5439 unsigned int pad_align = least_bit_hwi (pad) * BITS_PER_UNIT;
5440 parm_align = MIN (parm_align, pad_align);
5444 /* This isn't already where we want it on the stack, so put it there.
5445 This can either be done with push or copy insns. */
5446 if (!emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), NULL_RTX,
5447 parm_align, partial, reg, used - size, argblock,
5448 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space,
5449 ARGS_SIZE_RTX (arg->locate.alignment_pad), true))
5450 sibcall_failure = 1;
5452 /* Unless this is a partially-in-register argument, the argument is now
5453 in the stack. */
5454 if (partial == 0)
5455 arg->value = arg->stack;
5457 else
5459 /* BLKmode, at least partly to be pushed. */
5461 unsigned int parm_align;
5462 int excess;
5463 rtx size_rtx;
5465 /* Pushing a nonscalar.
5466 If part is passed in registers, PARTIAL says how much
5467 and emit_push_insn will take care of putting it there. */
5469 /* Round its size up to a multiple
5470 of the allocation unit for arguments. */
5472 if (arg->locate.size.var != 0)
5474 excess = 0;
5475 size_rtx = ARGS_SIZE_RTX (arg->locate.size);
5477 else
5479 /* PUSH_ROUNDING has no effect on us, because emit_push_insn
5480 for BLKmode is careful to avoid it. */
5481 excess = (arg->locate.size.constant
5482 - int_size_in_bytes (TREE_TYPE (pval))
5483 + partial);
5484 size_rtx = expand_expr (size_in_bytes (TREE_TYPE (pval)),
5485 NULL_RTX, TYPE_MODE (sizetype),
5486 EXPAND_NORMAL);
5489 parm_align = arg->locate.boundary;
5491 /* When an argument is padded down, the block is aligned to
5492 PARM_BOUNDARY, but the actual argument isn't. */
5493 if (FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)) == downward)
5495 if (arg->locate.size.var)
5496 parm_align = BITS_PER_UNIT;
5497 else if (excess)
5499 unsigned int excess_align = least_bit_hwi (excess) * BITS_PER_UNIT;
5500 parm_align = MIN (parm_align, excess_align);
5504 if ((flags & ECF_SIBCALL) && MEM_P (arg->value))
5506 /* emit_push_insn might not work properly if arg->value and
5507 argblock + arg->locate.offset areas overlap. */
5508 rtx x = arg->value;
5509 int i = 0;
5511 if (XEXP (x, 0) == crtl->args.internal_arg_pointer
5512 || (GET_CODE (XEXP (x, 0)) == PLUS
5513 && XEXP (XEXP (x, 0), 0) ==
5514 crtl->args.internal_arg_pointer
5515 && CONST_INT_P (XEXP (XEXP (x, 0), 1))))
5517 if (XEXP (x, 0) != crtl->args.internal_arg_pointer)
5518 i = INTVAL (XEXP (XEXP (x, 0), 1));
5520 /* arg.locate doesn't contain the pretend_args_size offset,
5521 it's part of argblock. Ensure we don't count it in I. */
5522 if (STACK_GROWS_DOWNWARD)
5523 i -= crtl->args.pretend_args_size;
5524 else
5525 i += crtl->args.pretend_args_size;
5527 /* expand_call should ensure this. */
5528 gcc_assert (!arg->locate.offset.var
5529 && arg->locate.size.var == 0
5530 && CONST_INT_P (size_rtx));
5532 if (arg->locate.offset.constant > i)
5534 if (arg->locate.offset.constant < i + INTVAL (size_rtx))
5535 sibcall_failure = 1;
5537 else if (arg->locate.offset.constant < i)
5539 /* Use arg->locate.size.constant instead of size_rtx
5540 because we only care about the part of the argument
5541 on the stack. */
5542 if (i < (arg->locate.offset.constant
5543 + arg->locate.size.constant))
5544 sibcall_failure = 1;
5546 else
5548 /* Even though they appear to be at the same location,
5549 if part of the outgoing argument is in registers,
5550 they aren't really at the same location. Check for
5551 this by making sure that the incoming size is the
5552 same as the outgoing size. */
5553 if (arg->locate.size.constant != INTVAL (size_rtx))
5554 sibcall_failure = 1;
5559 emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), size_rtx,
5560 parm_align, partial, reg, excess, argblock,
5561 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space,
5562 ARGS_SIZE_RTX (arg->locate.alignment_pad), false);
5564 /* Unless this is a partially-in-register argument, the argument is now
5565 in the stack.
5567 ??? Unlike the case above, in which we want the actual
5568 address of the data, so that we can load it directly into a
5569 register, here we want the address of the stack slot, so that
5570 it's properly aligned for word-by-word copying or something
5571 like that. It's not clear that this is always correct. */
5572 if (partial == 0)
5573 arg->value = arg->stack_slot;
5576 if (arg->reg && GET_CODE (arg->reg) == PARALLEL)
5578 tree type = TREE_TYPE (arg->tree_value);
5579 arg->parallel_value
5580 = emit_group_load_into_temps (arg->reg, arg->value, type,
5581 int_size_in_bytes (type));
5584 /* Mark all slots this store used. */
5585 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL)
5586 && argblock && ! variable_size && arg->stack)
5587 for (i = lower_bound; i < upper_bound; i++)
5588 stack_usage_map[i] = 1;
5590 /* Once we have pushed something, pops can't safely
5591 be deferred during the rest of the arguments. */
5592 NO_DEFER_POP;
5594 /* Free any temporary slots made in processing this argument. */
5595 pop_temp_slots ();
5597 return sibcall_failure;
5600 /* Nonzero if we do not know how to pass TYPE solely in registers. */
5602 bool
5603 must_pass_in_stack_var_size (machine_mode mode ATTRIBUTE_UNUSED,
5604 const_tree type)
5606 if (!type)
5607 return false;
5609 /* If the type has variable size... */
5610 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
5611 return true;
5613 /* If the type is marked as addressable (it is required
5614 to be constructed into the stack)... */
5615 if (TREE_ADDRESSABLE (type))
5616 return true;
5618 return false;
5621 /* Another version of the TARGET_MUST_PASS_IN_STACK hook. This one
5622 takes trailing padding of a structure into account. */
5623 /* ??? Should be able to merge these two by examining BLOCK_REG_PADDING. */
5625 bool
5626 must_pass_in_stack_var_size_or_pad (machine_mode mode, const_tree type)
5628 if (!type)
5629 return false;
5631 /* If the type has variable size... */
5632 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
5633 return true;
5635 /* If the type is marked as addressable (it is required
5636 to be constructed into the stack)... */
5637 if (TREE_ADDRESSABLE (type))
5638 return true;
5640 /* If the padding and mode of the type is such that a copy into
5641 a register would put it into the wrong part of the register. */
5642 if (mode == BLKmode
5643 && int_size_in_bytes (type) % (PARM_BOUNDARY / BITS_PER_UNIT)
5644 && (FUNCTION_ARG_PADDING (mode, type)
5645 == (BYTES_BIG_ENDIAN ? upward : downward)))
5646 return true;
5648 return false;
5651 /* Tell the garbage collector about GTY markers in this source file. */
5652 #include "gt-calls.h"