]>
Commit | Line | Data |
---|---|---|
1 | /* Convert function calls to rtl insns, for GNU C compiler. | |
2 | Copyright (C) 1989-2024 Free Software Foundation, Inc. | |
3 | ||
4 | This file is part of GCC. | |
5 | ||
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. | |
10 | ||
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. | |
15 | ||
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/>. */ | |
19 | ||
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-vrp.h" | |
51 | #include "tree-ssanames.h" | |
52 | #include "intl.h" | |
53 | #include "stringpool.h" | |
54 | #include "hash-map.h" | |
55 | #include "hash-traits.h" | |
56 | #include "attribs.h" | |
57 | #include "builtins.h" | |
58 | #include "gimple-iterator.h" | |
59 | #include "gimple-fold.h" | |
60 | #include "attr-fnspec.h" | |
61 | #include "value-query.h" | |
62 | #include "tree-pretty-print.h" | |
63 | #include "tree-eh.h" | |
64 | ||
65 | /* Like PREFERRED_STACK_BOUNDARY but in units of bytes, not bits. */ | |
66 | #define STACK_BYTES (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT) | |
67 | ||
68 | /* Data structure and subroutines used within expand_call. */ | |
69 | ||
70 | struct arg_data | |
71 | { | |
72 | /* Tree node for this argument. */ | |
73 | tree tree_value; | |
74 | /* Mode for value; TYPE_MODE unless promoted. */ | |
75 | machine_mode mode; | |
76 | /* Current RTL value for argument, or 0 if it isn't precomputed. */ | |
77 | rtx value; | |
78 | /* Initially-compute RTL value for argument; only for const functions. */ | |
79 | rtx initial_value; | |
80 | /* Register to pass this argument in, 0 if passed on stack, or an | |
81 | PARALLEL if the arg is to be copied into multiple non-contiguous | |
82 | registers. */ | |
83 | rtx reg; | |
84 | /* Register to pass this argument in when generating tail call sequence. | |
85 | This is not the same register as for normal calls on machines with | |
86 | register windows. */ | |
87 | rtx tail_call_reg; | |
88 | /* If REG is a PARALLEL, this is a copy of VALUE pulled into the correct | |
89 | form for emit_group_move. */ | |
90 | rtx parallel_value; | |
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 | /* True 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 false, just because FUNCTION_ARG says so. | |
100 | pass_on_stack identifies arguments that *cannot* go in registers. */ | |
101 | bool 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; | |
119 | }; | |
120 | ||
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; | |
126 | ||
127 | /* Size of STACK_USAGE_MAP. */ | |
128 | static unsigned int highest_outgoing_arg_in_use; | |
129 | ||
130 | /* Assume that any stack location at this byte index is used, | |
131 | without checking the contents of stack_usage_map. */ | |
132 | static unsigned HOST_WIDE_INT stack_usage_watermark = HOST_WIDE_INT_M1U; | |
133 | ||
134 | /* A bitmap of virtual-incoming stack space. Bit is set if the corresponding | |
135 | stack location's tail call argument has been already stored into the stack. | |
136 | This bitmap is used to prevent sibling call optimization if function tries | |
137 | to use parent's incoming argument slots when they have been already | |
138 | overwritten with tail call arguments. */ | |
139 | static sbitmap stored_args_map; | |
140 | ||
141 | /* Assume that any virtual-incoming location at this byte index has been | |
142 | stored, without checking the contents of stored_args_map. */ | |
143 | static unsigned HOST_WIDE_INT stored_args_watermark; | |
144 | ||
145 | /* stack_arg_under_construction is nonzero when an argument may be | |
146 | initialized with a constructor call (including a C function that | |
147 | returns a BLKmode struct) and expand_call must take special action | |
148 | to make sure the object being constructed does not overlap the | |
149 | argument list for the constructor call. */ | |
150 | static int stack_arg_under_construction; | |
151 | ||
152 | static void precompute_register_parameters (int, struct arg_data *, int *); | |
153 | static bool store_one_arg (struct arg_data *, rtx, int, int, int); | |
154 | static void store_unaligned_arguments_into_pseudos (struct arg_data *, int); | |
155 | static bool finalize_must_preallocate (bool, int, struct arg_data *, | |
156 | struct args_size *); | |
157 | static void precompute_arguments (int, struct arg_data *); | |
158 | static void compute_argument_addresses (struct arg_data *, rtx, int); | |
159 | static rtx rtx_for_function_call (tree, tree); | |
160 | static void load_register_parameters (struct arg_data *, int, rtx *, int, | |
161 | int, bool *); | |
162 | static int special_function_p (const_tree, int); | |
163 | static bool check_sibcall_argument_overlap_1 (rtx); | |
164 | static bool check_sibcall_argument_overlap (rtx_insn *, struct arg_data *, | |
165 | bool); | |
166 | static tree split_complex_types (tree); | |
167 | ||
168 | #ifdef REG_PARM_STACK_SPACE | |
169 | static rtx save_fixed_argument_area (int, rtx, int *, int *); | |
170 | static void restore_fixed_argument_area (rtx, rtx, int, int); | |
171 | #endif | |
172 | \f | |
173 | /* Return true if bytes [LOWER_BOUND, UPPER_BOUND) of the outgoing | |
174 | stack region might already be in use. */ | |
175 | ||
176 | static bool | |
177 | stack_region_maybe_used_p (poly_uint64 lower_bound, poly_uint64 upper_bound, | |
178 | unsigned int reg_parm_stack_space) | |
179 | { | |
180 | unsigned HOST_WIDE_INT const_lower, const_upper; | |
181 | const_lower = constant_lower_bound (lower_bound); | |
182 | if (!upper_bound.is_constant (&const_upper)) | |
183 | const_upper = HOST_WIDE_INT_M1U; | |
184 | ||
185 | if (const_upper > stack_usage_watermark) | |
186 | return true; | |
187 | ||
188 | /* Don't worry about things in the fixed argument area; | |
189 | it has already been saved. */ | |
190 | const_lower = MAX (const_lower, reg_parm_stack_space); | |
191 | const_upper = MIN (const_upper, highest_outgoing_arg_in_use); | |
192 | for (unsigned HOST_WIDE_INT i = const_lower; i < const_upper; ++i) | |
193 | if (stack_usage_map[i]) | |
194 | return true; | |
195 | return false; | |
196 | } | |
197 | ||
198 | /* Record that bytes [LOWER_BOUND, UPPER_BOUND) of the outgoing | |
199 | stack region are now in use. */ | |
200 | ||
201 | static void | |
202 | mark_stack_region_used (poly_uint64 lower_bound, poly_uint64 upper_bound) | |
203 | { | |
204 | unsigned HOST_WIDE_INT const_lower, const_upper; | |
205 | const_lower = constant_lower_bound (lower_bound); | |
206 | if (upper_bound.is_constant (&const_upper) | |
207 | && const_upper <= highest_outgoing_arg_in_use) | |
208 | for (unsigned HOST_WIDE_INT i = const_lower; i < const_upper; ++i) | |
209 | stack_usage_map[i] = 1; | |
210 | else | |
211 | stack_usage_watermark = MIN (stack_usage_watermark, const_lower); | |
212 | } | |
213 | ||
214 | /* Force FUNEXP into a form suitable for the address of a CALL, | |
215 | and return that as an rtx. Also load the static chain register | |
216 | if FNDECL is a nested function. | |
217 | ||
218 | CALL_FUSAGE points to a variable holding the prospective | |
219 | CALL_INSN_FUNCTION_USAGE information. */ | |
220 | ||
221 | rtx | |
222 | prepare_call_address (tree fndecl_or_type, rtx funexp, rtx static_chain_value, | |
223 | rtx *call_fusage, int reg_parm_seen, int flags) | |
224 | { | |
225 | /* Make a valid memory address and copy constants through pseudo-regs, | |
226 | but not for a constant address if -fno-function-cse. */ | |
227 | if (GET_CODE (funexp) != SYMBOL_REF) | |
228 | { | |
229 | /* If it's an indirect call by descriptor, generate code to perform | |
230 | runtime identification of the pointer and load the descriptor. */ | |
231 | if ((flags & ECF_BY_DESCRIPTOR) && !flag_trampolines) | |
232 | { | |
233 | const int bit_val = targetm.calls.custom_function_descriptors; | |
234 | rtx call_lab = gen_label_rtx (); | |
235 | ||
236 | gcc_assert (fndecl_or_type && TYPE_P (fndecl_or_type)); | |
237 | fndecl_or_type | |
238 | = build_decl (UNKNOWN_LOCATION, FUNCTION_DECL, NULL_TREE, | |
239 | fndecl_or_type); | |
240 | DECL_STATIC_CHAIN (fndecl_or_type) = 1; | |
241 | rtx chain = targetm.calls.static_chain (fndecl_or_type, false); | |
242 | ||
243 | if (GET_MODE (funexp) != Pmode) | |
244 | funexp = convert_memory_address (Pmode, funexp); | |
245 | ||
246 | /* Avoid long live ranges around function calls. */ | |
247 | funexp = copy_to_mode_reg (Pmode, funexp); | |
248 | ||
249 | if (REG_P (chain)) | |
250 | emit_insn (gen_rtx_CLOBBER (VOIDmode, chain)); | |
251 | ||
252 | /* Emit the runtime identification pattern. */ | |
253 | rtx mask = gen_rtx_AND (Pmode, funexp, GEN_INT (bit_val)); | |
254 | emit_cmp_and_jump_insns (mask, const0_rtx, EQ, NULL_RTX, Pmode, 1, | |
255 | call_lab); | |
256 | ||
257 | /* Statically predict the branch to very likely taken. */ | |
258 | rtx_insn *insn = get_last_insn (); | |
259 | if (JUMP_P (insn)) | |
260 | predict_insn_def (insn, PRED_BUILTIN_EXPECT, TAKEN); | |
261 | ||
262 | /* Load the descriptor. */ | |
263 | rtx mem = gen_rtx_MEM (ptr_mode, | |
264 | plus_constant (Pmode, funexp, - bit_val)); | |
265 | MEM_NOTRAP_P (mem) = 1; | |
266 | mem = convert_memory_address (Pmode, mem); | |
267 | emit_move_insn (chain, mem); | |
268 | ||
269 | mem = gen_rtx_MEM (ptr_mode, | |
270 | plus_constant (Pmode, funexp, | |
271 | POINTER_SIZE / BITS_PER_UNIT | |
272 | - bit_val)); | |
273 | MEM_NOTRAP_P (mem) = 1; | |
274 | mem = convert_memory_address (Pmode, mem); | |
275 | emit_move_insn (funexp, mem); | |
276 | ||
277 | emit_label (call_lab); | |
278 | ||
279 | if (REG_P (chain)) | |
280 | { | |
281 | use_reg (call_fusage, chain); | |
282 | STATIC_CHAIN_REG_P (chain) = 1; | |
283 | } | |
284 | ||
285 | /* Make sure we're not going to be overwritten below. */ | |
286 | gcc_assert (!static_chain_value); | |
287 | } | |
288 | ||
289 | /* If we are using registers for parameters, force the | |
290 | function address into a register now. */ | |
291 | funexp = ((reg_parm_seen | |
292 | && targetm.small_register_classes_for_mode_p (FUNCTION_MODE)) | |
293 | ? force_not_mem (memory_address (FUNCTION_MODE, funexp)) | |
294 | : memory_address (FUNCTION_MODE, funexp)); | |
295 | } | |
296 | else | |
297 | { | |
298 | /* funexp could be a SYMBOL_REF represents a function pointer which is | |
299 | of ptr_mode. In this case, it should be converted into address mode | |
300 | to be a valid address for memory rtx pattern. See PR 64971. */ | |
301 | if (GET_MODE (funexp) != Pmode) | |
302 | funexp = convert_memory_address (Pmode, funexp); | |
303 | ||
304 | if (!(flags & ECF_SIBCALL)) | |
305 | { | |
306 | if (!NO_FUNCTION_CSE && optimize && ! flag_no_function_cse) | |
307 | funexp = force_reg (Pmode, funexp); | |
308 | } | |
309 | } | |
310 | ||
311 | if (static_chain_value != 0 | |
312 | && (TREE_CODE (fndecl_or_type) != FUNCTION_DECL | |
313 | || DECL_STATIC_CHAIN (fndecl_or_type))) | |
314 | { | |
315 | rtx chain; | |
316 | ||
317 | chain = targetm.calls.static_chain (fndecl_or_type, false); | |
318 | static_chain_value = convert_memory_address (Pmode, static_chain_value); | |
319 | ||
320 | emit_move_insn (chain, static_chain_value); | |
321 | if (REG_P (chain)) | |
322 | { | |
323 | use_reg (call_fusage, chain); | |
324 | STATIC_CHAIN_REG_P (chain) = 1; | |
325 | } | |
326 | } | |
327 | ||
328 | return funexp; | |
329 | } | |
330 | ||
331 | /* Generate instructions to call function FUNEXP, | |
332 | and optionally pop the results. | |
333 | The CALL_INSN is the first insn generated. | |
334 | ||
335 | FNDECL is the declaration node of the function. This is given to the | |
336 | hook TARGET_RETURN_POPS_ARGS to determine whether this function pops | |
337 | its own args. | |
338 | ||
339 | FUNTYPE is the data type of the function. This is given to the hook | |
340 | TARGET_RETURN_POPS_ARGS to determine whether this function pops its | |
341 | own args. We used to allow an identifier for library functions, but | |
342 | that doesn't work when the return type is an aggregate type and the | |
343 | calling convention says that the pointer to this aggregate is to be | |
344 | popped by the callee. | |
345 | ||
346 | STACK_SIZE is the number of bytes of arguments on the stack, | |
347 | ROUNDED_STACK_SIZE is that number rounded up to | |
348 | PREFERRED_STACK_BOUNDARY; zero if the size is variable. This is | |
349 | both to put into the call insn and to generate explicit popping | |
350 | code if necessary. | |
351 | ||
352 | STRUCT_VALUE_SIZE is the number of bytes wanted in a structure value. | |
353 | It is zero if this call doesn't want a structure value. | |
354 | ||
355 | NEXT_ARG_REG is the rtx that results from executing | |
356 | targetm.calls.function_arg (&args_so_far, | |
357 | function_arg_info::end_marker ()); | |
358 | just after all the args have had their registers assigned. | |
359 | This could be whatever you like, but normally it is the first | |
360 | arg-register beyond those used for args in this call, | |
361 | or 0 if all the arg-registers are used in this call. | |
362 | It is passed on to `gen_call' so you can put this info in the call insn. | |
363 | ||
364 | VALREG is a hard register in which a value is returned, | |
365 | or 0 if the call does not return a value. | |
366 | ||
367 | OLD_INHIBIT_DEFER_POP is the value that `inhibit_defer_pop' had before | |
368 | the args to this call were processed. | |
369 | We restore `inhibit_defer_pop' to that value. | |
370 | ||
371 | CALL_FUSAGE is either empty or an EXPR_LIST of USE expressions that | |
372 | denote registers used by the called function. */ | |
373 | ||
374 | static void | |
375 | emit_call_1 (rtx funexp, tree fntree ATTRIBUTE_UNUSED, tree fndecl ATTRIBUTE_UNUSED, | |
376 | tree funtype ATTRIBUTE_UNUSED, | |
377 | poly_int64 stack_size ATTRIBUTE_UNUSED, | |
378 | poly_int64 rounded_stack_size, | |
379 | poly_int64 struct_value_size ATTRIBUTE_UNUSED, | |
380 | rtx next_arg_reg ATTRIBUTE_UNUSED, rtx valreg, | |
381 | int old_inhibit_defer_pop, rtx call_fusage, int ecf_flags, | |
382 | cumulative_args_t args_so_far ATTRIBUTE_UNUSED) | |
383 | { | |
384 | rtx rounded_stack_size_rtx = gen_int_mode (rounded_stack_size, Pmode); | |
385 | rtx call, funmem, pat; | |
386 | bool already_popped = false; | |
387 | poly_int64 n_popped = 0; | |
388 | ||
389 | /* Sibling call patterns never pop arguments (no sibcall(_value)_pop | |
390 | patterns exist). Any popping that the callee does on return will | |
391 | be from our caller's frame rather than ours. */ | |
392 | if (!(ecf_flags & ECF_SIBCALL)) | |
393 | { | |
394 | n_popped += targetm.calls.return_pops_args (fndecl, funtype, stack_size); | |
395 | ||
396 | #ifdef CALL_POPS_ARGS | |
397 | n_popped += CALL_POPS_ARGS (*get_cumulative_args (args_so_far)); | |
398 | #endif | |
399 | } | |
400 | ||
401 | /* Ensure address is valid. SYMBOL_REF is already valid, so no need, | |
402 | and we don't want to load it into a register as an optimization, | |
403 | because prepare_call_address already did it if it should be done. */ | |
404 | if (GET_CODE (funexp) != SYMBOL_REF) | |
405 | funexp = memory_address (FUNCTION_MODE, funexp); | |
406 | ||
407 | funmem = gen_rtx_MEM (FUNCTION_MODE, funexp); | |
408 | if (fndecl && TREE_CODE (fndecl) == FUNCTION_DECL) | |
409 | { | |
410 | tree t = fndecl; | |
411 | ||
412 | /* Although a built-in FUNCTION_DECL and its non-__builtin | |
413 | counterpart compare equal and get a shared mem_attrs, they | |
414 | produce different dump output in compare-debug compilations, | |
415 | if an entry gets garbage collected in one compilation, then | |
416 | adds a different (but equivalent) entry, while the other | |
417 | doesn't run the garbage collector at the same spot and then | |
418 | shares the mem_attr with the equivalent entry. */ | |
419 | if (DECL_BUILT_IN_CLASS (t) == BUILT_IN_NORMAL) | |
420 | { | |
421 | tree t2 = builtin_decl_explicit (DECL_FUNCTION_CODE (t)); | |
422 | if (t2) | |
423 | t = t2; | |
424 | } | |
425 | ||
426 | set_mem_expr (funmem, t); | |
427 | } | |
428 | else if (fntree) | |
429 | set_mem_expr (funmem, build_simple_mem_ref (CALL_EXPR_FN (fntree))); | |
430 | ||
431 | if (ecf_flags & ECF_SIBCALL) | |
432 | { | |
433 | if (valreg) | |
434 | pat = targetm.gen_sibcall_value (valreg, funmem, | |
435 | rounded_stack_size_rtx, | |
436 | next_arg_reg, NULL_RTX); | |
437 | else | |
438 | pat = targetm.gen_sibcall (funmem, rounded_stack_size_rtx, | |
439 | next_arg_reg, | |
440 | gen_int_mode (struct_value_size, Pmode)); | |
441 | } | |
442 | /* If the target has "call" or "call_value" insns, then prefer them | |
443 | if no arguments are actually popped. If the target does not have | |
444 | "call" or "call_value" insns, then we must use the popping versions | |
445 | even if the call has no arguments to pop. */ | |
446 | else if (maybe_ne (n_popped, 0) | |
447 | || !(valreg | |
448 | ? targetm.have_call_value () | |
449 | : targetm.have_call ())) | |
450 | { | |
451 | rtx n_pop = gen_int_mode (n_popped, Pmode); | |
452 | ||
453 | /* If this subroutine pops its own args, record that in the call insn | |
454 | if possible, for the sake of frame pointer elimination. */ | |
455 | ||
456 | if (valreg) | |
457 | pat = targetm.gen_call_value_pop (valreg, funmem, | |
458 | rounded_stack_size_rtx, | |
459 | next_arg_reg, n_pop); | |
460 | else | |
461 | pat = targetm.gen_call_pop (funmem, rounded_stack_size_rtx, | |
462 | next_arg_reg, n_pop); | |
463 | ||
464 | already_popped = true; | |
465 | } | |
466 | else | |
467 | { | |
468 | if (valreg) | |
469 | pat = targetm.gen_call_value (valreg, funmem, rounded_stack_size_rtx, | |
470 | next_arg_reg, NULL_RTX); | |
471 | else | |
472 | pat = targetm.gen_call (funmem, rounded_stack_size_rtx, next_arg_reg, | |
473 | gen_int_mode (struct_value_size, Pmode)); | |
474 | } | |
475 | emit_insn (pat); | |
476 | ||
477 | /* Find the call we just emitted. */ | |
478 | rtx_call_insn *call_insn = last_call_insn (); | |
479 | ||
480 | /* Some target create a fresh MEM instead of reusing the one provided | |
481 | above. Set its MEM_EXPR. */ | |
482 | call = get_call_rtx_from (call_insn); | |
483 | if (call | |
484 | && MEM_EXPR (XEXP (call, 0)) == NULL_TREE | |
485 | && MEM_EXPR (funmem) != NULL_TREE) | |
486 | set_mem_expr (XEXP (call, 0), MEM_EXPR (funmem)); | |
487 | ||
488 | /* Put the register usage information there. */ | |
489 | add_function_usage_to (call_insn, call_fusage); | |
490 | ||
491 | /* If this is a const call, then set the insn's unchanging bit. */ | |
492 | if (ecf_flags & ECF_CONST) | |
493 | RTL_CONST_CALL_P (call_insn) = 1; | |
494 | ||
495 | /* If this is a pure call, then set the insn's unchanging bit. */ | |
496 | if (ecf_flags & ECF_PURE) | |
497 | RTL_PURE_CALL_P (call_insn) = 1; | |
498 | ||
499 | /* If this is a const call, then set the insn's unchanging bit. */ | |
500 | if (ecf_flags & ECF_LOOPING_CONST_OR_PURE) | |
501 | RTL_LOOPING_CONST_OR_PURE_CALL_P (call_insn) = 1; | |
502 | ||
503 | /* Create a nothrow REG_EH_REGION note, if needed. */ | |
504 | make_reg_eh_region_note (call_insn, ecf_flags, 0); | |
505 | ||
506 | if (ecf_flags & ECF_NORETURN) | |
507 | add_reg_note (call_insn, REG_NORETURN, const0_rtx); | |
508 | ||
509 | if (ecf_flags & ECF_RETURNS_TWICE) | |
510 | { | |
511 | add_reg_note (call_insn, REG_SETJMP, const0_rtx); | |
512 | cfun->calls_setjmp = 1; | |
513 | } | |
514 | ||
515 | SIBLING_CALL_P (call_insn) = ((ecf_flags & ECF_SIBCALL) != 0); | |
516 | ||
517 | /* Restore this now, so that we do defer pops for this call's args | |
518 | if the context of the call as a whole permits. */ | |
519 | inhibit_defer_pop = old_inhibit_defer_pop; | |
520 | ||
521 | if (maybe_ne (n_popped, 0)) | |
522 | { | |
523 | if (!already_popped) | |
524 | CALL_INSN_FUNCTION_USAGE (call_insn) | |
525 | = gen_rtx_EXPR_LIST (VOIDmode, | |
526 | gen_rtx_CLOBBER (VOIDmode, stack_pointer_rtx), | |
527 | CALL_INSN_FUNCTION_USAGE (call_insn)); | |
528 | rounded_stack_size -= n_popped; | |
529 | rounded_stack_size_rtx = gen_int_mode (rounded_stack_size, Pmode); | |
530 | stack_pointer_delta -= n_popped; | |
531 | ||
532 | add_args_size_note (call_insn, stack_pointer_delta); | |
533 | ||
534 | /* If popup is needed, stack realign must use DRAP */ | |
535 | if (SUPPORTS_STACK_ALIGNMENT) | |
536 | crtl->need_drap = true; | |
537 | } | |
538 | /* For noreturn calls when not accumulating outgoing args force | |
539 | REG_ARGS_SIZE note to prevent crossjumping of calls with different | |
540 | args sizes. */ | |
541 | else if (!ACCUMULATE_OUTGOING_ARGS && (ecf_flags & ECF_NORETURN) != 0) | |
542 | add_args_size_note (call_insn, stack_pointer_delta); | |
543 | ||
544 | if (!ACCUMULATE_OUTGOING_ARGS) | |
545 | { | |
546 | /* If returning from the subroutine does not automatically pop the args, | |
547 | we need an instruction to pop them sooner or later. | |
548 | Perhaps do it now; perhaps just record how much space to pop later. | |
549 | ||
550 | If returning from the subroutine does pop the args, indicate that the | |
551 | stack pointer will be changed. */ | |
552 | ||
553 | if (maybe_ne (rounded_stack_size, 0)) | |
554 | { | |
555 | if (ecf_flags & ECF_NORETURN) | |
556 | /* Just pretend we did the pop. */ | |
557 | stack_pointer_delta -= rounded_stack_size; | |
558 | else if (flag_defer_pop && inhibit_defer_pop == 0 | |
559 | && ! (ecf_flags & (ECF_CONST | ECF_PURE))) | |
560 | pending_stack_adjust += rounded_stack_size; | |
561 | else | |
562 | adjust_stack (rounded_stack_size_rtx); | |
563 | } | |
564 | } | |
565 | /* When we accumulate outgoing args, we must avoid any stack manipulations. | |
566 | Restore the stack pointer to its original value now. Usually | |
567 | ACCUMULATE_OUTGOING_ARGS targets don't get here, but there are exceptions. | |
568 | On i386 ACCUMULATE_OUTGOING_ARGS can be enabled on demand, and | |
569 | popping variants of functions exist as well. | |
570 | ||
571 | ??? We may optimize similar to defer_pop above, but it is | |
572 | probably not worthwhile. | |
573 | ||
574 | ??? It will be worthwhile to enable combine_stack_adjustments even for | |
575 | such machines. */ | |
576 | else if (maybe_ne (n_popped, 0)) | |
577 | anti_adjust_stack (gen_int_mode (n_popped, Pmode)); | |
578 | } | |
579 | ||
580 | /* Determine if the function identified by FNDECL is one with | |
581 | special properties we wish to know about. Modify FLAGS accordingly. | |
582 | ||
583 | For example, if the function might return more than one time (setjmp), then | |
584 | set ECF_RETURNS_TWICE. | |
585 | ||
586 | Set ECF_MAY_BE_ALLOCA for any memory allocation function that might allocate | |
587 | space from the stack such as alloca. */ | |
588 | ||
589 | static int | |
590 | special_function_p (const_tree fndecl, int flags) | |
591 | { | |
592 | tree name_decl = DECL_NAME (fndecl); | |
593 | ||
594 | if (maybe_special_function_p (fndecl) | |
595 | && IDENTIFIER_LENGTH (name_decl) <= 11) | |
596 | { | |
597 | const char *name = IDENTIFIER_POINTER (name_decl); | |
598 | const char *tname = name; | |
599 | ||
600 | /* We assume that alloca will always be called by name. It | |
601 | makes no sense to pass it as a pointer-to-function to | |
602 | anything that does not understand its behavior. */ | |
603 | if (IDENTIFIER_LENGTH (name_decl) == 6 | |
604 | && name[0] == 'a' | |
605 | && ! strcmp (name, "alloca")) | |
606 | flags |= ECF_MAY_BE_ALLOCA; | |
607 | ||
608 | /* Disregard prefix _ or __. */ | |
609 | if (name[0] == '_') | |
610 | { | |
611 | if (name[1] == '_') | |
612 | tname += 2; | |
613 | else | |
614 | tname += 1; | |
615 | } | |
616 | ||
617 | /* ECF_RETURNS_TWICE is safe even for -ffreestanding. */ | |
618 | if (! strcmp (tname, "setjmp") | |
619 | || ! strcmp (tname, "sigsetjmp") | |
620 | || ! strcmp (name, "savectx") | |
621 | || ! strcmp (name, "vfork") | |
622 | || ! strcmp (name, "getcontext")) | |
623 | flags |= ECF_RETURNS_TWICE; | |
624 | } | |
625 | ||
626 | if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL | |
627 | && ALLOCA_FUNCTION_CODE_P (DECL_FUNCTION_CODE (fndecl))) | |
628 | flags |= ECF_MAY_BE_ALLOCA; | |
629 | ||
630 | return flags; | |
631 | } | |
632 | ||
633 | /* Return fnspec for DECL. */ | |
634 | ||
635 | static attr_fnspec | |
636 | decl_fnspec (tree fndecl) | |
637 | { | |
638 | tree attr; | |
639 | tree type = TREE_TYPE (fndecl); | |
640 | if (type) | |
641 | { | |
642 | attr = lookup_attribute ("fn spec", TYPE_ATTRIBUTES (type)); | |
643 | if (attr) | |
644 | { | |
645 | return TREE_VALUE (TREE_VALUE (attr)); | |
646 | } | |
647 | } | |
648 | if (fndecl_built_in_p (fndecl, BUILT_IN_NORMAL)) | |
649 | return builtin_fnspec (fndecl); | |
650 | return ""; | |
651 | } | |
652 | ||
653 | /* Similar to special_function_p; return a set of ERF_ flags for the | |
654 | function FNDECL. */ | |
655 | static int | |
656 | decl_return_flags (tree fndecl) | |
657 | { | |
658 | attr_fnspec fnspec = decl_fnspec (fndecl); | |
659 | ||
660 | unsigned int arg; | |
661 | if (fnspec.returns_arg (&arg)) | |
662 | return ERF_RETURNS_ARG | arg; | |
663 | ||
664 | if (fnspec.returns_noalias_p ()) | |
665 | return ERF_NOALIAS; | |
666 | return 0; | |
667 | } | |
668 | ||
669 | /* Return true when FNDECL represents a call to setjmp. */ | |
670 | ||
671 | bool | |
672 | setjmp_call_p (const_tree fndecl) | |
673 | { | |
674 | if (DECL_IS_RETURNS_TWICE (fndecl)) | |
675 | return true; | |
676 | if (special_function_p (fndecl, 0) & ECF_RETURNS_TWICE) | |
677 | return true; | |
678 | ||
679 | return false; | |
680 | } | |
681 | ||
682 | ||
683 | /* Return true if STMT may be an alloca call. */ | |
684 | ||
685 | bool | |
686 | gimple_maybe_alloca_call_p (const gimple *stmt) | |
687 | { | |
688 | tree fndecl; | |
689 | ||
690 | if (!is_gimple_call (stmt)) | |
691 | return false; | |
692 | ||
693 | fndecl = gimple_call_fndecl (stmt); | |
694 | if (fndecl && (special_function_p (fndecl, 0) & ECF_MAY_BE_ALLOCA)) | |
695 | return true; | |
696 | ||
697 | return false; | |
698 | } | |
699 | ||
700 | /* Return true if STMT is a builtin alloca call. */ | |
701 | ||
702 | bool | |
703 | gimple_alloca_call_p (const gimple *stmt) | |
704 | { | |
705 | tree fndecl; | |
706 | ||
707 | if (!is_gimple_call (stmt)) | |
708 | return false; | |
709 | ||
710 | fndecl = gimple_call_fndecl (stmt); | |
711 | if (fndecl && fndecl_built_in_p (fndecl, BUILT_IN_NORMAL)) | |
712 | switch (DECL_FUNCTION_CODE (fndecl)) | |
713 | { | |
714 | CASE_BUILT_IN_ALLOCA: | |
715 | return gimple_call_num_args (stmt) > 0; | |
716 | default: | |
717 | break; | |
718 | } | |
719 | ||
720 | return false; | |
721 | } | |
722 | ||
723 | /* Return true when exp contains a builtin alloca call. */ | |
724 | ||
725 | bool | |
726 | alloca_call_p (const_tree exp) | |
727 | { | |
728 | tree fndecl; | |
729 | if (TREE_CODE (exp) == CALL_EXPR | |
730 | && (fndecl = get_callee_fndecl (exp)) | |
731 | && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
732 | switch (DECL_FUNCTION_CODE (fndecl)) | |
733 | { | |
734 | CASE_BUILT_IN_ALLOCA: | |
735 | return true; | |
736 | default: | |
737 | break; | |
738 | } | |
739 | ||
740 | return false; | |
741 | } | |
742 | ||
743 | /* Return TRUE if FNDECL is either a TM builtin or a TM cloned | |
744 | function. Return FALSE otherwise. */ | |
745 | ||
746 | static bool | |
747 | is_tm_builtin (const_tree fndecl) | |
748 | { | |
749 | if (fndecl == NULL) | |
750 | return false; | |
751 | ||
752 | if (decl_is_tm_clone (fndecl)) | |
753 | return true; | |
754 | ||
755 | if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
756 | { | |
757 | switch (DECL_FUNCTION_CODE (fndecl)) | |
758 | { | |
759 | case BUILT_IN_TM_COMMIT: | |
760 | case BUILT_IN_TM_COMMIT_EH: | |
761 | case BUILT_IN_TM_ABORT: | |
762 | case BUILT_IN_TM_IRREVOCABLE: | |
763 | case BUILT_IN_TM_GETTMCLONE_IRR: | |
764 | case BUILT_IN_TM_MEMCPY: | |
765 | case BUILT_IN_TM_MEMMOVE: | |
766 | case BUILT_IN_TM_MEMSET: | |
767 | CASE_BUILT_IN_TM_STORE (1): | |
768 | CASE_BUILT_IN_TM_STORE (2): | |
769 | CASE_BUILT_IN_TM_STORE (4): | |
770 | CASE_BUILT_IN_TM_STORE (8): | |
771 | CASE_BUILT_IN_TM_STORE (FLOAT): | |
772 | CASE_BUILT_IN_TM_STORE (DOUBLE): | |
773 | CASE_BUILT_IN_TM_STORE (LDOUBLE): | |
774 | CASE_BUILT_IN_TM_STORE (M64): | |
775 | CASE_BUILT_IN_TM_STORE (M128): | |
776 | CASE_BUILT_IN_TM_STORE (M256): | |
777 | CASE_BUILT_IN_TM_LOAD (1): | |
778 | CASE_BUILT_IN_TM_LOAD (2): | |
779 | CASE_BUILT_IN_TM_LOAD (4): | |
780 | CASE_BUILT_IN_TM_LOAD (8): | |
781 | CASE_BUILT_IN_TM_LOAD (FLOAT): | |
782 | CASE_BUILT_IN_TM_LOAD (DOUBLE): | |
783 | CASE_BUILT_IN_TM_LOAD (LDOUBLE): | |
784 | CASE_BUILT_IN_TM_LOAD (M64): | |
785 | CASE_BUILT_IN_TM_LOAD (M128): | |
786 | CASE_BUILT_IN_TM_LOAD (M256): | |
787 | case BUILT_IN_TM_LOG: | |
788 | case BUILT_IN_TM_LOG_1: | |
789 | case BUILT_IN_TM_LOG_2: | |
790 | case BUILT_IN_TM_LOG_4: | |
791 | case BUILT_IN_TM_LOG_8: | |
792 | case BUILT_IN_TM_LOG_FLOAT: | |
793 | case BUILT_IN_TM_LOG_DOUBLE: | |
794 | case BUILT_IN_TM_LOG_LDOUBLE: | |
795 | case BUILT_IN_TM_LOG_M64: | |
796 | case BUILT_IN_TM_LOG_M128: | |
797 | case BUILT_IN_TM_LOG_M256: | |
798 | return true; | |
799 | default: | |
800 | break; | |
801 | } | |
802 | } | |
803 | return false; | |
804 | } | |
805 | ||
806 | /* Detect flags (function attributes) from the function decl or type node. */ | |
807 | ||
808 | int | |
809 | flags_from_decl_or_type (const_tree exp) | |
810 | { | |
811 | int flags = 0; | |
812 | ||
813 | if (DECL_P (exp)) | |
814 | { | |
815 | /* The function exp may have the `malloc' attribute. */ | |
816 | if (DECL_IS_MALLOC (exp)) | |
817 | flags |= ECF_MALLOC; | |
818 | ||
819 | /* The function exp may have the `returns_twice' attribute. */ | |
820 | if (DECL_IS_RETURNS_TWICE (exp)) | |
821 | flags |= ECF_RETURNS_TWICE; | |
822 | ||
823 | /* Process the pure and const attributes. */ | |
824 | if (TREE_READONLY (exp)) | |
825 | flags |= ECF_CONST; | |
826 | if (DECL_PURE_P (exp)) | |
827 | flags |= ECF_PURE; | |
828 | if (DECL_LOOPING_CONST_OR_PURE_P (exp)) | |
829 | flags |= ECF_LOOPING_CONST_OR_PURE; | |
830 | ||
831 | if (DECL_IS_NOVOPS (exp)) | |
832 | flags |= ECF_NOVOPS; | |
833 | if (lookup_attribute ("leaf", DECL_ATTRIBUTES (exp))) | |
834 | flags |= ECF_LEAF; | |
835 | if (lookup_attribute ("cold", DECL_ATTRIBUTES (exp))) | |
836 | flags |= ECF_COLD; | |
837 | ||
838 | if (TREE_NOTHROW (exp)) | |
839 | flags |= ECF_NOTHROW; | |
840 | ||
841 | if (flag_tm) | |
842 | { | |
843 | if (is_tm_builtin (exp)) | |
844 | flags |= ECF_TM_BUILTIN; | |
845 | else if ((flags & (ECF_CONST|ECF_NOVOPS)) != 0 | |
846 | || lookup_attribute ("transaction_pure", | |
847 | TYPE_ATTRIBUTES (TREE_TYPE (exp)))) | |
848 | flags |= ECF_TM_PURE; | |
849 | } | |
850 | ||
851 | if (lookup_attribute ("expected_throw", DECL_ATTRIBUTES (exp))) | |
852 | flags |= ECF_XTHROW; | |
853 | ||
854 | flags = special_function_p (exp, flags); | |
855 | } | |
856 | else if (TYPE_P (exp)) | |
857 | { | |
858 | if (TYPE_READONLY (exp)) | |
859 | flags |= ECF_CONST; | |
860 | ||
861 | if (flag_tm | |
862 | && ((flags & ECF_CONST) != 0 | |
863 | || lookup_attribute ("transaction_pure", TYPE_ATTRIBUTES (exp)))) | |
864 | flags |= ECF_TM_PURE; | |
865 | } | |
866 | else | |
867 | gcc_unreachable (); | |
868 | ||
869 | if (TREE_THIS_VOLATILE (exp)) | |
870 | { | |
871 | flags |= ECF_NORETURN; | |
872 | if (flags & (ECF_CONST|ECF_PURE)) | |
873 | flags |= ECF_LOOPING_CONST_OR_PURE; | |
874 | } | |
875 | ||
876 | return flags; | |
877 | } | |
878 | ||
879 | /* Detect flags from a CALL_EXPR. */ | |
880 | ||
881 | int | |
882 | call_expr_flags (const_tree t) | |
883 | { | |
884 | int flags; | |
885 | tree decl = get_callee_fndecl (t); | |
886 | ||
887 | if (decl) | |
888 | flags = flags_from_decl_or_type (decl); | |
889 | else if (CALL_EXPR_FN (t) == NULL_TREE) | |
890 | flags = internal_fn_flags (CALL_EXPR_IFN (t)); | |
891 | else | |
892 | { | |
893 | tree type = TREE_TYPE (CALL_EXPR_FN (t)); | |
894 | if (type && TREE_CODE (type) == POINTER_TYPE) | |
895 | flags = flags_from_decl_or_type (TREE_TYPE (type)); | |
896 | else | |
897 | flags = 0; | |
898 | if (CALL_EXPR_BY_DESCRIPTOR (t)) | |
899 | flags |= ECF_BY_DESCRIPTOR; | |
900 | } | |
901 | ||
902 | return flags; | |
903 | } | |
904 | ||
905 | /* Return true if ARG should be passed by invisible reference. */ | |
906 | ||
907 | bool | |
908 | pass_by_reference (CUMULATIVE_ARGS *ca, function_arg_info arg) | |
909 | { | |
910 | if (tree type = arg.type) | |
911 | { | |
912 | /* If this type contains non-trivial constructors, then it is | |
913 | forbidden for the middle-end to create any new copies. */ | |
914 | if (TREE_ADDRESSABLE (type)) | |
915 | return true; | |
916 | ||
917 | /* GCC post 3.4 passes *all* variable sized types by reference. */ | |
918 | if (!TYPE_SIZE (type) || !poly_int_tree_p (TYPE_SIZE (type))) | |
919 | return true; | |
920 | ||
921 | /* If a record type should be passed the same as its first (and only) | |
922 | member, use the type and mode of that member. */ | |
923 | if (TREE_CODE (type) == RECORD_TYPE && TYPE_TRANSPARENT_AGGR (type)) | |
924 | { | |
925 | arg.type = TREE_TYPE (first_field (type)); | |
926 | arg.mode = TYPE_MODE (arg.type); | |
927 | } | |
928 | } | |
929 | ||
930 | return targetm.calls.pass_by_reference (pack_cumulative_args (ca), arg); | |
931 | } | |
932 | ||
933 | /* Return true if TYPE should be passed by reference when passed to | |
934 | the "..." arguments of a function. */ | |
935 | ||
936 | bool | |
937 | pass_va_arg_by_reference (tree type) | |
938 | { | |
939 | return pass_by_reference (NULL, function_arg_info (type, /*named=*/false)); | |
940 | } | |
941 | ||
942 | /* Decide whether ARG, which occurs in the state described by CA, | |
943 | should be passed by reference. Return true if so and update | |
944 | ARG accordingly. */ | |
945 | ||
946 | bool | |
947 | apply_pass_by_reference_rules (CUMULATIVE_ARGS *ca, function_arg_info &arg) | |
948 | { | |
949 | if (pass_by_reference (ca, arg)) | |
950 | { | |
951 | arg.type = build_pointer_type (arg.type); | |
952 | arg.mode = TYPE_MODE (arg.type); | |
953 | arg.pass_by_reference = true; | |
954 | return true; | |
955 | } | |
956 | return false; | |
957 | } | |
958 | ||
959 | /* Return true if ARG, which is passed by reference, should be callee | |
960 | copied instead of caller copied. */ | |
961 | ||
962 | bool | |
963 | reference_callee_copied (CUMULATIVE_ARGS *ca, const function_arg_info &arg) | |
964 | { | |
965 | if (arg.type && TREE_ADDRESSABLE (arg.type)) | |
966 | return false; | |
967 | return targetm.calls.callee_copies (pack_cumulative_args (ca), arg); | |
968 | } | |
969 | ||
970 | ||
971 | /* Precompute all register parameters as described by ARGS, storing values | |
972 | into fields within the ARGS array. | |
973 | ||
974 | NUM_ACTUALS indicates the total number elements in the ARGS array. | |
975 | ||
976 | Set REG_PARM_SEEN if we encounter a register parameter. */ | |
977 | ||
978 | static void | |
979 | precompute_register_parameters (int num_actuals, struct arg_data *args, | |
980 | int *reg_parm_seen) | |
981 | { | |
982 | int i; | |
983 | ||
984 | *reg_parm_seen = 0; | |
985 | ||
986 | for (i = 0; i < num_actuals; i++) | |
987 | if (args[i].reg != 0 && ! args[i].pass_on_stack) | |
988 | { | |
989 | *reg_parm_seen = 1; | |
990 | ||
991 | if (args[i].value == 0) | |
992 | { | |
993 | push_temp_slots (); | |
994 | args[i].value = expand_normal (args[i].tree_value); | |
995 | preserve_temp_slots (args[i].value); | |
996 | pop_temp_slots (); | |
997 | } | |
998 | ||
999 | /* If we are to promote the function arg to a wider mode, | |
1000 | do it now. */ | |
1001 | ||
1002 | machine_mode old_mode = TYPE_MODE (TREE_TYPE (args[i].tree_value)); | |
1003 | ||
1004 | /* Some ABIs require scalar floating point modes to be returned | |
1005 | in a wider scalar integer mode. We need to explicitly | |
1006 | reinterpret to an integer mode of the correct precision | |
1007 | before extending to the desired result. */ | |
1008 | if (SCALAR_INT_MODE_P (args[i].mode) | |
1009 | && SCALAR_FLOAT_MODE_P (old_mode) | |
1010 | && known_gt (GET_MODE_SIZE (args[i].mode), | |
1011 | GET_MODE_SIZE (old_mode))) | |
1012 | args[i].value = convert_float_to_wider_int (args[i].mode, old_mode, | |
1013 | args[i].value); | |
1014 | else if (args[i].mode != old_mode) | |
1015 | args[i].value = convert_modes (args[i].mode, old_mode, | |
1016 | args[i].value, args[i].unsignedp); | |
1017 | ||
1018 | /* If the value is a non-legitimate constant, force it into a | |
1019 | pseudo now. TLS symbols sometimes need a call to resolve. */ | |
1020 | if (CONSTANT_P (args[i].value) | |
1021 | && (!targetm.legitimate_constant_p (args[i].mode, args[i].value) | |
1022 | || targetm.precompute_tls_p (args[i].mode, args[i].value))) | |
1023 | args[i].value = force_reg (args[i].mode, args[i].value); | |
1024 | ||
1025 | /* If we're going to have to load the value by parts, pull the | |
1026 | parts into pseudos. The part extraction process can involve | |
1027 | non-trivial computation. */ | |
1028 | if (GET_CODE (args[i].reg) == PARALLEL) | |
1029 | { | |
1030 | tree type = TREE_TYPE (args[i].tree_value); | |
1031 | args[i].parallel_value | |
1032 | = emit_group_load_into_temps (args[i].reg, args[i].value, | |
1033 | type, int_size_in_bytes (type)); | |
1034 | } | |
1035 | ||
1036 | /* If the value is expensive, and we are inside an appropriately | |
1037 | short loop, put the value into a pseudo and then put the pseudo | |
1038 | into the hard reg. | |
1039 | ||
1040 | For small register classes, also do this if this call uses | |
1041 | register parameters. This is to avoid reload conflicts while | |
1042 | loading the parameters registers. */ | |
1043 | ||
1044 | else if ((! (REG_P (args[i].value) | |
1045 | || (GET_CODE (args[i].value) == SUBREG | |
1046 | && REG_P (SUBREG_REG (args[i].value))))) | |
1047 | && args[i].mode != BLKmode | |
1048 | && (set_src_cost (args[i].value, args[i].mode, | |
1049 | optimize_insn_for_speed_p ()) | |
1050 | > COSTS_N_INSNS (1)) | |
1051 | && ((*reg_parm_seen | |
1052 | && targetm.small_register_classes_for_mode_p (args[i].mode)) | |
1053 | || optimize)) | |
1054 | args[i].value = copy_to_mode_reg (args[i].mode, args[i].value); | |
1055 | } | |
1056 | } | |
1057 | ||
1058 | #ifdef REG_PARM_STACK_SPACE | |
1059 | ||
1060 | /* The argument list is the property of the called routine and it | |
1061 | may clobber it. If the fixed area has been used for previous | |
1062 | parameters, we must save and restore it. */ | |
1063 | ||
1064 | static rtx | |
1065 | save_fixed_argument_area (int reg_parm_stack_space, rtx argblock, int *low_to_save, int *high_to_save) | |
1066 | { | |
1067 | unsigned int low; | |
1068 | unsigned int high; | |
1069 | ||
1070 | /* Compute the boundary of the area that needs to be saved, if any. */ | |
1071 | high = reg_parm_stack_space; | |
1072 | if (ARGS_GROW_DOWNWARD) | |
1073 | high += 1; | |
1074 | ||
1075 | if (high > highest_outgoing_arg_in_use) | |
1076 | high = highest_outgoing_arg_in_use; | |
1077 | ||
1078 | for (low = 0; low < high; low++) | |
1079 | if (stack_usage_map[low] != 0 || low >= stack_usage_watermark) | |
1080 | { | |
1081 | int num_to_save; | |
1082 | machine_mode save_mode; | |
1083 | int delta; | |
1084 | rtx addr; | |
1085 | rtx stack_area; | |
1086 | rtx save_area; | |
1087 | ||
1088 | while (stack_usage_map[--high] == 0) | |
1089 | ; | |
1090 | ||
1091 | *low_to_save = low; | |
1092 | *high_to_save = high; | |
1093 | ||
1094 | num_to_save = high - low + 1; | |
1095 | ||
1096 | /* If we don't have the required alignment, must do this | |
1097 | in BLKmode. */ | |
1098 | scalar_int_mode imode; | |
1099 | if (int_mode_for_size (num_to_save * BITS_PER_UNIT, 1).exists (&imode) | |
1100 | && (low & (MIN (GET_MODE_SIZE (imode), | |
1101 | BIGGEST_ALIGNMENT / UNITS_PER_WORD) - 1)) == 0) | |
1102 | save_mode = imode; | |
1103 | else | |
1104 | save_mode = BLKmode; | |
1105 | ||
1106 | if (ARGS_GROW_DOWNWARD) | |
1107 | delta = -high; | |
1108 | else | |
1109 | delta = low; | |
1110 | ||
1111 | addr = plus_constant (Pmode, argblock, delta); | |
1112 | stack_area = gen_rtx_MEM (save_mode, memory_address (save_mode, addr)); | |
1113 | ||
1114 | set_mem_align (stack_area, PARM_BOUNDARY); | |
1115 | if (save_mode == BLKmode) | |
1116 | { | |
1117 | save_area = assign_stack_temp (BLKmode, num_to_save); | |
1118 | emit_block_move (validize_mem (save_area), stack_area, | |
1119 | GEN_INT (num_to_save), BLOCK_OP_CALL_PARM); | |
1120 | } | |
1121 | else | |
1122 | { | |
1123 | save_area = gen_reg_rtx (save_mode); | |
1124 | emit_move_insn (save_area, stack_area); | |
1125 | } | |
1126 | ||
1127 | return save_area; | |
1128 | } | |
1129 | ||
1130 | return NULL_RTX; | |
1131 | } | |
1132 | ||
1133 | static void | |
1134 | restore_fixed_argument_area (rtx save_area, rtx argblock, int high_to_save, int low_to_save) | |
1135 | { | |
1136 | machine_mode save_mode = GET_MODE (save_area); | |
1137 | int delta; | |
1138 | rtx addr, stack_area; | |
1139 | ||
1140 | if (ARGS_GROW_DOWNWARD) | |
1141 | delta = -high_to_save; | |
1142 | else | |
1143 | delta = low_to_save; | |
1144 | ||
1145 | addr = plus_constant (Pmode, argblock, delta); | |
1146 | stack_area = gen_rtx_MEM (save_mode, memory_address (save_mode, addr)); | |
1147 | set_mem_align (stack_area, PARM_BOUNDARY); | |
1148 | ||
1149 | if (save_mode != BLKmode) | |
1150 | emit_move_insn (stack_area, save_area); | |
1151 | else | |
1152 | emit_block_move (stack_area, validize_mem (save_area), | |
1153 | GEN_INT (high_to_save - low_to_save + 1), | |
1154 | BLOCK_OP_CALL_PARM); | |
1155 | } | |
1156 | #endif /* REG_PARM_STACK_SPACE */ | |
1157 | ||
1158 | /* If any elements in ARGS refer to parameters that are to be passed in | |
1159 | registers, but not in memory, and whose alignment does not permit a | |
1160 | direct copy into registers. Copy the values into a group of pseudos | |
1161 | which we will later copy into the appropriate hard registers. | |
1162 | ||
1163 | Pseudos for each unaligned argument will be stored into the array | |
1164 | args[argnum].aligned_regs. The caller is responsible for deallocating | |
1165 | the aligned_regs array if it is nonzero. */ | |
1166 | ||
1167 | static void | |
1168 | store_unaligned_arguments_into_pseudos (struct arg_data *args, int num_actuals) | |
1169 | { | |
1170 | int i, j; | |
1171 | ||
1172 | for (i = 0; i < num_actuals; i++) | |
1173 | if (args[i].reg != 0 && ! args[i].pass_on_stack | |
1174 | && GET_CODE (args[i].reg) != PARALLEL | |
1175 | && args[i].mode == BLKmode | |
1176 | && MEM_P (args[i].value) | |
1177 | && (MEM_ALIGN (args[i].value) | |
1178 | < (unsigned int) MIN (BIGGEST_ALIGNMENT, BITS_PER_WORD))) | |
1179 | { | |
1180 | int bytes = int_size_in_bytes (TREE_TYPE (args[i].tree_value)); | |
1181 | int endian_correction = 0; | |
1182 | ||
1183 | if (args[i].partial) | |
1184 | { | |
1185 | gcc_assert (args[i].partial % UNITS_PER_WORD == 0); | |
1186 | args[i].n_aligned_regs = args[i].partial / UNITS_PER_WORD; | |
1187 | } | |
1188 | else | |
1189 | { | |
1190 | args[i].n_aligned_regs | |
1191 | = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
1192 | } | |
1193 | ||
1194 | args[i].aligned_regs = XNEWVEC (rtx, args[i].n_aligned_regs); | |
1195 | ||
1196 | /* Structures smaller than a word are normally aligned to the | |
1197 | least significant byte. On a BYTES_BIG_ENDIAN machine, | |
1198 | this means we must skip the empty high order bytes when | |
1199 | calculating the bit offset. */ | |
1200 | if (bytes < UNITS_PER_WORD | |
1201 | #ifdef BLOCK_REG_PADDING | |
1202 | && (BLOCK_REG_PADDING (args[i].mode, | |
1203 | TREE_TYPE (args[i].tree_value), 1) | |
1204 | == PAD_DOWNWARD) | |
1205 | #else | |
1206 | && BYTES_BIG_ENDIAN | |
1207 | #endif | |
1208 | ) | |
1209 | endian_correction = BITS_PER_WORD - bytes * BITS_PER_UNIT; | |
1210 | ||
1211 | for (j = 0; j < args[i].n_aligned_regs; j++) | |
1212 | { | |
1213 | rtx reg = gen_reg_rtx (word_mode); | |
1214 | rtx word = operand_subword_force (args[i].value, j, BLKmode); | |
1215 | int bitsize = MIN (bytes * BITS_PER_UNIT, BITS_PER_WORD); | |
1216 | ||
1217 | args[i].aligned_regs[j] = reg; | |
1218 | word = extract_bit_field (word, bitsize, 0, 1, NULL_RTX, | |
1219 | word_mode, word_mode, false, NULL); | |
1220 | ||
1221 | /* There is no need to restrict this code to loading items | |
1222 | in TYPE_ALIGN sized hunks. The bitfield instructions can | |
1223 | load up entire word sized registers efficiently. | |
1224 | ||
1225 | ??? This may not be needed anymore. | |
1226 | We use to emit a clobber here but that doesn't let later | |
1227 | passes optimize the instructions we emit. By storing 0 into | |
1228 | the register later passes know the first AND to zero out the | |
1229 | bitfield being set in the register is unnecessary. The store | |
1230 | of 0 will be deleted as will at least the first AND. */ | |
1231 | ||
1232 | emit_move_insn (reg, const0_rtx); | |
1233 | ||
1234 | bytes -= bitsize / BITS_PER_UNIT; | |
1235 | store_bit_field (reg, bitsize, endian_correction, 0, 0, | |
1236 | word_mode, word, false, false); | |
1237 | } | |
1238 | } | |
1239 | } | |
1240 | ||
1241 | /* Issue an error if CALL_EXPR was flagged as requiring | |
1242 | tall-call optimization. */ | |
1243 | ||
1244 | void | |
1245 | maybe_complain_about_tail_call (tree call_expr, const char *reason) | |
1246 | { | |
1247 | gcc_assert (TREE_CODE (call_expr) == CALL_EXPR); | |
1248 | if (!CALL_EXPR_MUST_TAIL_CALL (call_expr)) | |
1249 | return; | |
1250 | ||
1251 | error_at (EXPR_LOCATION (call_expr), "cannot tail-call: %s", reason); | |
1252 | } | |
1253 | ||
1254 | /* Fill in ARGS_SIZE and ARGS array based on the parameters found in | |
1255 | CALL_EXPR EXP. | |
1256 | ||
1257 | NUM_ACTUALS is the total number of parameters. | |
1258 | ||
1259 | N_NAMED_ARGS is the total number of named arguments. | |
1260 | ||
1261 | STRUCT_VALUE_ADDR_VALUE is the implicit argument for a struct return | |
1262 | value, or null. | |
1263 | ||
1264 | FNDECL is the tree code for the target of this call (if known) | |
1265 | ||
1266 | ARGS_SO_FAR holds state needed by the target to know where to place | |
1267 | the next argument. | |
1268 | ||
1269 | REG_PARM_STACK_SPACE is the number of bytes of stack space reserved | |
1270 | for arguments which are passed in registers. | |
1271 | ||
1272 | OLD_STACK_LEVEL is a pointer to an rtx which olds the old stack level | |
1273 | and may be modified by this routine. | |
1274 | ||
1275 | OLD_PENDING_ADJ and FLAGS are pointers to integer flags which | |
1276 | may be modified by this routine. | |
1277 | ||
1278 | MUST_PREALLOCATE is a pointer to bool which may be | |
1279 | modified by this routine. | |
1280 | ||
1281 | MAY_TAILCALL is cleared if we encounter an invisible pass-by-reference | |
1282 | that requires allocation of stack space. | |
1283 | ||
1284 | CALL_FROM_THUNK_P is true if this call is the jump from a thunk to | |
1285 | the thunked-to function. */ | |
1286 | ||
1287 | static void | |
1288 | initialize_argument_information (int num_actuals ATTRIBUTE_UNUSED, | |
1289 | struct arg_data *args, | |
1290 | struct args_size *args_size, | |
1291 | int n_named_args ATTRIBUTE_UNUSED, | |
1292 | tree exp, tree struct_value_addr_value, | |
1293 | tree fndecl, tree fntype, | |
1294 | cumulative_args_t args_so_far, | |
1295 | int reg_parm_stack_space, | |
1296 | rtx *old_stack_level, | |
1297 | poly_int64 *old_pending_adj, | |
1298 | bool *must_preallocate, int *ecf_flags, | |
1299 | bool *may_tailcall, bool call_from_thunk_p) | |
1300 | { | |
1301 | CUMULATIVE_ARGS *args_so_far_pnt = get_cumulative_args (args_so_far); | |
1302 | location_t loc = EXPR_LOCATION (exp); | |
1303 | ||
1304 | /* Count arg position in order args appear. */ | |
1305 | int argpos; | |
1306 | ||
1307 | int i; | |
1308 | ||
1309 | args_size->constant = 0; | |
1310 | args_size->var = 0; | |
1311 | ||
1312 | /* In this loop, we consider args in the order they are written. | |
1313 | We fill up ARGS from the back. */ | |
1314 | ||
1315 | i = num_actuals - 1; | |
1316 | { | |
1317 | int j = i; | |
1318 | call_expr_arg_iterator iter; | |
1319 | tree arg; | |
1320 | ||
1321 | if (struct_value_addr_value) | |
1322 | { | |
1323 | args[j].tree_value = struct_value_addr_value; | |
1324 | j--; | |
1325 | } | |
1326 | argpos = 0; | |
1327 | FOR_EACH_CALL_EXPR_ARG (arg, iter, exp) | |
1328 | { | |
1329 | tree argtype = TREE_TYPE (arg); | |
1330 | ||
1331 | if (targetm.calls.split_complex_arg | |
1332 | && argtype | |
1333 | && TREE_CODE (argtype) == COMPLEX_TYPE | |
1334 | && targetm.calls.split_complex_arg (argtype)) | |
1335 | { | |
1336 | tree subtype = TREE_TYPE (argtype); | |
1337 | args[j].tree_value = build1 (REALPART_EXPR, subtype, arg); | |
1338 | j--; | |
1339 | args[j].tree_value = build1 (IMAGPART_EXPR, subtype, arg); | |
1340 | } | |
1341 | else | |
1342 | args[j].tree_value = arg; | |
1343 | j--; | |
1344 | argpos++; | |
1345 | } | |
1346 | } | |
1347 | ||
1348 | /* I counts args in order (to be) pushed; ARGPOS counts in order written. */ | |
1349 | for (argpos = 0; argpos < num_actuals; i--, argpos++) | |
1350 | { | |
1351 | tree type = TREE_TYPE (args[i].tree_value); | |
1352 | int unsignedp; | |
1353 | ||
1354 | /* Replace erroneous argument with constant zero. */ | |
1355 | if (type == error_mark_node || !COMPLETE_TYPE_P (type)) | |
1356 | args[i].tree_value = integer_zero_node, type = integer_type_node; | |
1357 | ||
1358 | /* If TYPE is a transparent union or record, pass things the way | |
1359 | we would pass the first field of the union or record. We have | |
1360 | already verified that the modes are the same. */ | |
1361 | if (RECORD_OR_UNION_TYPE_P (type) && TYPE_TRANSPARENT_AGGR (type)) | |
1362 | type = TREE_TYPE (first_field (type)); | |
1363 | ||
1364 | /* Decide where to pass this arg. | |
1365 | ||
1366 | args[i].reg is nonzero if all or part is passed in registers. | |
1367 | ||
1368 | args[i].partial is nonzero if part but not all is passed in registers, | |
1369 | and the exact value says how many bytes are passed in registers. | |
1370 | ||
1371 | args[i].pass_on_stack is true if the argument must at least be | |
1372 | computed on the stack. It may then be loaded back into registers | |
1373 | if args[i].reg is nonzero. | |
1374 | ||
1375 | These decisions are driven by the FUNCTION_... macros and must agree | |
1376 | with those made by function.cc. */ | |
1377 | ||
1378 | /* See if this argument should be passed by invisible reference. */ | |
1379 | function_arg_info arg (type, argpos < n_named_args); | |
1380 | if (pass_by_reference (args_so_far_pnt, arg)) | |
1381 | { | |
1382 | const bool callee_copies | |
1383 | = reference_callee_copied (args_so_far_pnt, arg); | |
1384 | tree base; | |
1385 | ||
1386 | /* If we're compiling a thunk, pass directly the address of an object | |
1387 | already in memory, instead of making a copy. Likewise if we want | |
1388 | to make the copy in the callee instead of the caller. */ | |
1389 | if ((call_from_thunk_p || callee_copies) | |
1390 | && TREE_CODE (args[i].tree_value) != WITH_SIZE_EXPR | |
1391 | && ((base = get_base_address (args[i].tree_value)), true) | |
1392 | && TREE_CODE (base) != SSA_NAME | |
1393 | && (!DECL_P (base) || MEM_P (DECL_RTL (base)))) | |
1394 | { | |
1395 | /* We may have turned the parameter value into an SSA name. | |
1396 | Go back to the original parameter so we can take the | |
1397 | address. */ | |
1398 | if (TREE_CODE (args[i].tree_value) == SSA_NAME) | |
1399 | { | |
1400 | gcc_assert (SSA_NAME_IS_DEFAULT_DEF (args[i].tree_value)); | |
1401 | args[i].tree_value = SSA_NAME_VAR (args[i].tree_value); | |
1402 | gcc_assert (TREE_CODE (args[i].tree_value) == PARM_DECL); | |
1403 | } | |
1404 | /* Argument setup code may have copied the value to register. We | |
1405 | revert that optimization now because the tail call code must | |
1406 | use the original location. */ | |
1407 | if (TREE_CODE (args[i].tree_value) == PARM_DECL | |
1408 | && !MEM_P (DECL_RTL (args[i].tree_value)) | |
1409 | && DECL_INCOMING_RTL (args[i].tree_value) | |
1410 | && MEM_P (DECL_INCOMING_RTL (args[i].tree_value))) | |
1411 | set_decl_rtl (args[i].tree_value, | |
1412 | DECL_INCOMING_RTL (args[i].tree_value)); | |
1413 | ||
1414 | mark_addressable (args[i].tree_value); | |
1415 | ||
1416 | /* We can't use sibcalls if a callee-copied argument is | |
1417 | stored in the current function's frame. */ | |
1418 | if (!call_from_thunk_p && DECL_P (base) && !TREE_STATIC (base)) | |
1419 | { | |
1420 | *may_tailcall = false; | |
1421 | maybe_complain_about_tail_call (exp, | |
1422 | "a callee-copied argument is" | |
1423 | " stored in the current" | |
1424 | " function's frame"); | |
1425 | } | |
1426 | ||
1427 | args[i].tree_value = build_fold_addr_expr_loc (loc, | |
1428 | args[i].tree_value); | |
1429 | type = TREE_TYPE (args[i].tree_value); | |
1430 | ||
1431 | if (*ecf_flags & ECF_CONST) | |
1432 | *ecf_flags &= ~(ECF_CONST | ECF_LOOPING_CONST_OR_PURE); | |
1433 | } | |
1434 | else | |
1435 | { | |
1436 | /* We make a copy of the object and pass the address to the | |
1437 | function being called. */ | |
1438 | rtx copy; | |
1439 | ||
1440 | if (!COMPLETE_TYPE_P (type) | |
1441 | || TREE_CODE (TYPE_SIZE_UNIT (type)) != INTEGER_CST | |
1442 | || (flag_stack_check == GENERIC_STACK_CHECK | |
1443 | && compare_tree_int (TYPE_SIZE_UNIT (type), | |
1444 | STACK_CHECK_MAX_VAR_SIZE) > 0)) | |
1445 | { | |
1446 | /* This is a variable-sized object. Make space on the stack | |
1447 | for it. */ | |
1448 | rtx size_rtx = expr_size (args[i].tree_value); | |
1449 | ||
1450 | if (*old_stack_level == 0) | |
1451 | { | |
1452 | emit_stack_save (SAVE_BLOCK, old_stack_level); | |
1453 | *old_pending_adj = pending_stack_adjust; | |
1454 | pending_stack_adjust = 0; | |
1455 | } | |
1456 | ||
1457 | /* We can pass TRUE as the 4th argument because we just | |
1458 | saved the stack pointer and will restore it right after | |
1459 | the call. */ | |
1460 | copy = allocate_dynamic_stack_space (size_rtx, | |
1461 | TYPE_ALIGN (type), | |
1462 | TYPE_ALIGN (type), | |
1463 | max_int_size_in_bytes | |
1464 | (type), | |
1465 | true); | |
1466 | copy = gen_rtx_MEM (BLKmode, copy); | |
1467 | set_mem_attributes (copy, type, 1); | |
1468 | } | |
1469 | else | |
1470 | copy = assign_temp (type, 1, 0); | |
1471 | ||
1472 | store_expr (args[i].tree_value, copy, 0, false, false); | |
1473 | ||
1474 | /* Just change the const function to pure and then let | |
1475 | the next test clear the pure based on | |
1476 | callee_copies. */ | |
1477 | if (*ecf_flags & ECF_CONST) | |
1478 | { | |
1479 | *ecf_flags &= ~ECF_CONST; | |
1480 | *ecf_flags |= ECF_PURE; | |
1481 | } | |
1482 | ||
1483 | if (!callee_copies && *ecf_flags & ECF_PURE) | |
1484 | *ecf_flags &= ~(ECF_PURE | ECF_LOOPING_CONST_OR_PURE); | |
1485 | ||
1486 | args[i].tree_value | |
1487 | = build_fold_addr_expr_loc (loc, make_tree (type, copy)); | |
1488 | type = TREE_TYPE (args[i].tree_value); | |
1489 | *may_tailcall = false; | |
1490 | maybe_complain_about_tail_call (exp, | |
1491 | "argument must be passed" | |
1492 | " by copying"); | |
1493 | } | |
1494 | arg.pass_by_reference = true; | |
1495 | } | |
1496 | ||
1497 | unsignedp = TYPE_UNSIGNED (type); | |
1498 | arg.type = type; | |
1499 | arg.mode | |
1500 | = promote_function_mode (type, TYPE_MODE (type), &unsignedp, | |
1501 | fndecl ? TREE_TYPE (fndecl) : fntype, 0); | |
1502 | ||
1503 | args[i].unsignedp = unsignedp; | |
1504 | args[i].mode = arg.mode; | |
1505 | ||
1506 | targetm.calls.warn_parameter_passing_abi (args_so_far, type); | |
1507 | ||
1508 | args[i].reg = targetm.calls.function_arg (args_so_far, arg); | |
1509 | ||
1510 | /* If this is a sibling call and the machine has register windows, the | |
1511 | register window has to be unwinded before calling the routine, so | |
1512 | arguments have to go into the incoming registers. */ | |
1513 | if (targetm.calls.function_incoming_arg != targetm.calls.function_arg) | |
1514 | args[i].tail_call_reg | |
1515 | = targetm.calls.function_incoming_arg (args_so_far, arg); | |
1516 | else | |
1517 | args[i].tail_call_reg = args[i].reg; | |
1518 | ||
1519 | if (args[i].reg) | |
1520 | args[i].partial = targetm.calls.arg_partial_bytes (args_so_far, arg); | |
1521 | ||
1522 | args[i].pass_on_stack = targetm.calls.must_pass_in_stack (arg); | |
1523 | ||
1524 | /* If FUNCTION_ARG returned a (parallel [(expr_list (nil) ...) ...]), | |
1525 | it means that we are to pass this arg in the register(s) designated | |
1526 | by the PARALLEL, but also to pass it in the stack. */ | |
1527 | if (args[i].reg && GET_CODE (args[i].reg) == PARALLEL | |
1528 | && XEXP (XVECEXP (args[i].reg, 0, 0), 0) == 0) | |
1529 | args[i].pass_on_stack = true; | |
1530 | ||
1531 | /* If this is an addressable type, we must preallocate the stack | |
1532 | since we must evaluate the object into its final location. | |
1533 | ||
1534 | If this is to be passed in both registers and the stack, it is simpler | |
1535 | to preallocate. */ | |
1536 | if (TREE_ADDRESSABLE (type) | |
1537 | || (args[i].pass_on_stack && args[i].reg != 0)) | |
1538 | *must_preallocate = true; | |
1539 | ||
1540 | /* Compute the stack-size of this argument. */ | |
1541 | if (args[i].reg == 0 || args[i].partial != 0 | |
1542 | || reg_parm_stack_space > 0 | |
1543 | || args[i].pass_on_stack) | |
1544 | locate_and_pad_parm (arg.mode, type, | |
1545 | #ifdef STACK_PARMS_IN_REG_PARM_AREA | |
1546 | 1, | |
1547 | #else | |
1548 | args[i].reg != 0, | |
1549 | #endif | |
1550 | reg_parm_stack_space, | |
1551 | args[i].pass_on_stack ? 0 : args[i].partial, | |
1552 | fndecl, args_size, &args[i].locate); | |
1553 | #ifdef BLOCK_REG_PADDING | |
1554 | else | |
1555 | /* The argument is passed entirely in registers. See at which | |
1556 | end it should be padded. */ | |
1557 | args[i].locate.where_pad = | |
1558 | BLOCK_REG_PADDING (arg.mode, type, | |
1559 | int_size_in_bytes (type) <= UNITS_PER_WORD); | |
1560 | #endif | |
1561 | ||
1562 | /* Update ARGS_SIZE, the total stack space for args so far. */ | |
1563 | ||
1564 | args_size->constant += args[i].locate.size.constant; | |
1565 | if (args[i].locate.size.var) | |
1566 | ADD_PARM_SIZE (*args_size, args[i].locate.size.var); | |
1567 | ||
1568 | /* Increment ARGS_SO_FAR, which has info about which arg-registers | |
1569 | have been used, etc. */ | |
1570 | ||
1571 | /* ??? Traditionally we've passed TYPE_MODE here, instead of the | |
1572 | promoted_mode used for function_arg above. However, the | |
1573 | corresponding handling of incoming arguments in function.cc | |
1574 | does pass the promoted mode. */ | |
1575 | arg.mode = TYPE_MODE (type); | |
1576 | targetm.calls.function_arg_advance (args_so_far, arg); | |
1577 | } | |
1578 | } | |
1579 | ||
1580 | /* Update ARGS_SIZE to contain the total size for the argument block. | |
1581 | Return the original constant component of the argument block's size. | |
1582 | ||
1583 | REG_PARM_STACK_SPACE holds the number of bytes of stack space reserved | |
1584 | for arguments passed in registers. */ | |
1585 | ||
1586 | static poly_int64 | |
1587 | compute_argument_block_size (int reg_parm_stack_space, | |
1588 | struct args_size *args_size, | |
1589 | tree fndecl ATTRIBUTE_UNUSED, | |
1590 | tree fntype ATTRIBUTE_UNUSED, | |
1591 | int preferred_stack_boundary ATTRIBUTE_UNUSED) | |
1592 | { | |
1593 | poly_int64 unadjusted_args_size = args_size->constant; | |
1594 | ||
1595 | /* For accumulate outgoing args mode we don't need to align, since the frame | |
1596 | will be already aligned. Align to STACK_BOUNDARY in order to prevent | |
1597 | backends from generating misaligned frame sizes. */ | |
1598 | if (ACCUMULATE_OUTGOING_ARGS && preferred_stack_boundary > STACK_BOUNDARY) | |
1599 | preferred_stack_boundary = STACK_BOUNDARY; | |
1600 | ||
1601 | /* Compute the actual size of the argument block required. The variable | |
1602 | and constant sizes must be combined, the size may have to be rounded, | |
1603 | and there may be a minimum required size. */ | |
1604 | ||
1605 | if (args_size->var) | |
1606 | { | |
1607 | args_size->var = ARGS_SIZE_TREE (*args_size); | |
1608 | args_size->constant = 0; | |
1609 | ||
1610 | preferred_stack_boundary /= BITS_PER_UNIT; | |
1611 | if (preferred_stack_boundary > 1) | |
1612 | { | |
1613 | /* We don't handle this case yet. To handle it correctly we have | |
1614 | to add the delta, round and subtract the delta. | |
1615 | Currently no machine description requires this support. */ | |
1616 | gcc_assert (multiple_p (stack_pointer_delta, | |
1617 | preferred_stack_boundary)); | |
1618 | args_size->var = round_up (args_size->var, preferred_stack_boundary); | |
1619 | } | |
1620 | ||
1621 | if (reg_parm_stack_space > 0) | |
1622 | { | |
1623 | args_size->var | |
1624 | = size_binop (MAX_EXPR, args_size->var, | |
1625 | ssize_int (reg_parm_stack_space)); | |
1626 | ||
1627 | /* The area corresponding to register parameters is not to count in | |
1628 | the size of the block we need. So make the adjustment. */ | |
1629 | if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))) | |
1630 | args_size->var | |
1631 | = size_binop (MINUS_EXPR, args_size->var, | |
1632 | ssize_int (reg_parm_stack_space)); | |
1633 | } | |
1634 | } | |
1635 | else | |
1636 | { | |
1637 | preferred_stack_boundary /= BITS_PER_UNIT; | |
1638 | if (preferred_stack_boundary < 1) | |
1639 | preferred_stack_boundary = 1; | |
1640 | args_size->constant = (aligned_upper_bound (args_size->constant | |
1641 | + stack_pointer_delta, | |
1642 | preferred_stack_boundary) | |
1643 | - stack_pointer_delta); | |
1644 | ||
1645 | args_size->constant = upper_bound (args_size->constant, | |
1646 | reg_parm_stack_space); | |
1647 | ||
1648 | if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))) | |
1649 | args_size->constant -= reg_parm_stack_space; | |
1650 | } | |
1651 | return unadjusted_args_size; | |
1652 | } | |
1653 | ||
1654 | /* Precompute parameters as needed for a function call. | |
1655 | ||
1656 | FLAGS is mask of ECF_* constants. | |
1657 | ||
1658 | NUM_ACTUALS is the number of arguments. | |
1659 | ||
1660 | ARGS is an array containing information for each argument; this | |
1661 | routine fills in the INITIAL_VALUE and VALUE fields for each | |
1662 | precomputed argument. */ | |
1663 | ||
1664 | static void | |
1665 | precompute_arguments (int num_actuals, struct arg_data *args) | |
1666 | { | |
1667 | int i; | |
1668 | ||
1669 | /* If this is a libcall, then precompute all arguments so that we do not | |
1670 | get extraneous instructions emitted as part of the libcall sequence. */ | |
1671 | ||
1672 | /* If we preallocated the stack space, and some arguments must be passed | |
1673 | on the stack, then we must precompute any parameter which contains a | |
1674 | function call which will store arguments on the stack. | |
1675 | Otherwise, evaluating the parameter may clobber previous parameters | |
1676 | which have already been stored into the stack. (we have code to avoid | |
1677 | such case by saving the outgoing stack arguments, but it results in | |
1678 | worse code) */ | |
1679 | if (!ACCUMULATE_OUTGOING_ARGS) | |
1680 | return; | |
1681 | ||
1682 | for (i = 0; i < num_actuals; i++) | |
1683 | { | |
1684 | tree type; | |
1685 | machine_mode mode; | |
1686 | ||
1687 | if (TREE_CODE (args[i].tree_value) != CALL_EXPR) | |
1688 | continue; | |
1689 | ||
1690 | /* If this is an addressable type, we cannot pre-evaluate it. */ | |
1691 | type = TREE_TYPE (args[i].tree_value); | |
1692 | gcc_assert (!TREE_ADDRESSABLE (type)); | |
1693 | ||
1694 | args[i].initial_value = args[i].value | |
1695 | = expand_normal (args[i].tree_value); | |
1696 | ||
1697 | mode = TYPE_MODE (type); | |
1698 | if (mode != args[i].mode) | |
1699 | { | |
1700 | int unsignedp = args[i].unsignedp; | |
1701 | args[i].value | |
1702 | = convert_modes (args[i].mode, mode, | |
1703 | args[i].value, args[i].unsignedp); | |
1704 | ||
1705 | /* CSE will replace this only if it contains args[i].value | |
1706 | pseudo, so convert it down to the declared mode using | |
1707 | a SUBREG. */ | |
1708 | if (REG_P (args[i].value) | |
1709 | && GET_MODE_CLASS (args[i].mode) == MODE_INT | |
1710 | && promote_mode (type, mode, &unsignedp) != args[i].mode) | |
1711 | { | |
1712 | args[i].initial_value | |
1713 | = gen_lowpart_SUBREG (mode, args[i].value); | |
1714 | SUBREG_PROMOTED_VAR_P (args[i].initial_value) = 1; | |
1715 | SUBREG_PROMOTED_SET (args[i].initial_value, args[i].unsignedp); | |
1716 | } | |
1717 | } | |
1718 | } | |
1719 | } | |
1720 | ||
1721 | /* Given the current state of MUST_PREALLOCATE and information about | |
1722 | arguments to a function call in NUM_ACTUALS, ARGS and ARGS_SIZE, | |
1723 | compute and return the final value for MUST_PREALLOCATE. */ | |
1724 | ||
1725 | static bool | |
1726 | finalize_must_preallocate (bool must_preallocate, int num_actuals, | |
1727 | struct arg_data *args, struct args_size *args_size) | |
1728 | { | |
1729 | /* See if we have or want to preallocate stack space. | |
1730 | ||
1731 | If we would have to push a partially-in-regs parm | |
1732 | before other stack parms, preallocate stack space instead. | |
1733 | ||
1734 | If the size of some parm is not a multiple of the required stack | |
1735 | alignment, we must preallocate. | |
1736 | ||
1737 | If the total size of arguments that would otherwise create a copy in | |
1738 | a temporary (such as a CALL) is more than half the total argument list | |
1739 | size, preallocation is faster. | |
1740 | ||
1741 | Another reason to preallocate is if we have a machine (like the m88k) | |
1742 | where stack alignment is required to be maintained between every | |
1743 | pair of insns, not just when the call is made. However, we assume here | |
1744 | that such machines either do not have push insns (and hence preallocation | |
1745 | would occur anyway) or the problem is taken care of with | |
1746 | PUSH_ROUNDING. */ | |
1747 | ||
1748 | if (! must_preallocate) | |
1749 | { | |
1750 | bool partial_seen = false; | |
1751 | poly_int64 copy_to_evaluate_size = 0; | |
1752 | int i; | |
1753 | ||
1754 | for (i = 0; i < num_actuals && ! must_preallocate; i++) | |
1755 | { | |
1756 | if (args[i].partial > 0 && ! args[i].pass_on_stack) | |
1757 | partial_seen = true; | |
1758 | else if (partial_seen && args[i].reg == 0) | |
1759 | must_preallocate = true; | |
1760 | ||
1761 | if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode | |
1762 | && (TREE_CODE (args[i].tree_value) == CALL_EXPR | |
1763 | || TREE_CODE (args[i].tree_value) == TARGET_EXPR | |
1764 | || TREE_CODE (args[i].tree_value) == COND_EXPR | |
1765 | || TREE_ADDRESSABLE (TREE_TYPE (args[i].tree_value)))) | |
1766 | copy_to_evaluate_size | |
1767 | += int_size_in_bytes (TREE_TYPE (args[i].tree_value)); | |
1768 | } | |
1769 | ||
1770 | if (maybe_ne (args_size->constant, 0) | |
1771 | && maybe_ge (copy_to_evaluate_size * 2, args_size->constant)) | |
1772 | must_preallocate = true; | |
1773 | } | |
1774 | return must_preallocate; | |
1775 | } | |
1776 | ||
1777 | /* If we preallocated stack space, compute the address of each argument | |
1778 | and store it into the ARGS array. | |
1779 | ||
1780 | We need not ensure it is a valid memory address here; it will be | |
1781 | validized when it is used. | |
1782 | ||
1783 | ARGBLOCK is an rtx for the address of the outgoing arguments. */ | |
1784 | ||
1785 | static void | |
1786 | compute_argument_addresses (struct arg_data *args, rtx argblock, int num_actuals) | |
1787 | { | |
1788 | if (argblock) | |
1789 | { | |
1790 | rtx arg_reg = argblock; | |
1791 | int i; | |
1792 | poly_int64 arg_offset = 0; | |
1793 | ||
1794 | if (GET_CODE (argblock) == PLUS) | |
1795 | { | |
1796 | arg_reg = XEXP (argblock, 0); | |
1797 | arg_offset = rtx_to_poly_int64 (XEXP (argblock, 1)); | |
1798 | } | |
1799 | ||
1800 | for (i = 0; i < num_actuals; i++) | |
1801 | { | |
1802 | rtx offset = ARGS_SIZE_RTX (args[i].locate.offset); | |
1803 | rtx slot_offset = ARGS_SIZE_RTX (args[i].locate.slot_offset); | |
1804 | rtx addr; | |
1805 | unsigned int align, boundary; | |
1806 | poly_uint64 units_on_stack = 0; | |
1807 | machine_mode partial_mode = VOIDmode; | |
1808 | ||
1809 | /* Skip this parm if it will not be passed on the stack. */ | |
1810 | if (! args[i].pass_on_stack | |
1811 | && args[i].reg != 0 | |
1812 | && args[i].partial == 0) | |
1813 | continue; | |
1814 | ||
1815 | if (TYPE_EMPTY_P (TREE_TYPE (args[i].tree_value))) | |
1816 | continue; | |
1817 | ||
1818 | addr = simplify_gen_binary (PLUS, Pmode, arg_reg, offset); | |
1819 | addr = plus_constant (Pmode, addr, arg_offset); | |
1820 | ||
1821 | if (args[i].partial != 0) | |
1822 | { | |
1823 | /* Only part of the parameter is being passed on the stack. | |
1824 | Generate a simple memory reference of the correct size. */ | |
1825 | units_on_stack = args[i].locate.size.constant; | |
1826 | poly_uint64 bits_on_stack = units_on_stack * BITS_PER_UNIT; | |
1827 | partial_mode = int_mode_for_size (bits_on_stack, 1).else_blk (); | |
1828 | args[i].stack = gen_rtx_MEM (partial_mode, addr); | |
1829 | set_mem_size (args[i].stack, units_on_stack); | |
1830 | } | |
1831 | else | |
1832 | { | |
1833 | args[i].stack = gen_rtx_MEM (args[i].mode, addr); | |
1834 | set_mem_attributes (args[i].stack, | |
1835 | TREE_TYPE (args[i].tree_value), 1); | |
1836 | } | |
1837 | align = BITS_PER_UNIT; | |
1838 | boundary = args[i].locate.boundary; | |
1839 | poly_int64 offset_val; | |
1840 | if (args[i].locate.where_pad != PAD_DOWNWARD) | |
1841 | align = boundary; | |
1842 | else if (poly_int_rtx_p (offset, &offset_val)) | |
1843 | { | |
1844 | align = least_bit_hwi (boundary); | |
1845 | unsigned int offset_align | |
1846 | = known_alignment (offset_val) * BITS_PER_UNIT; | |
1847 | if (offset_align != 0) | |
1848 | align = MIN (align, offset_align); | |
1849 | } | |
1850 | set_mem_align (args[i].stack, align); | |
1851 | ||
1852 | addr = simplify_gen_binary (PLUS, Pmode, arg_reg, slot_offset); | |
1853 | addr = plus_constant (Pmode, addr, arg_offset); | |
1854 | ||
1855 | if (args[i].partial != 0) | |
1856 | { | |
1857 | /* Only part of the parameter is being passed on the stack. | |
1858 | Generate a simple memory reference of the correct size. | |
1859 | */ | |
1860 | args[i].stack_slot = gen_rtx_MEM (partial_mode, addr); | |
1861 | set_mem_size (args[i].stack_slot, units_on_stack); | |
1862 | } | |
1863 | else | |
1864 | { | |
1865 | args[i].stack_slot = gen_rtx_MEM (args[i].mode, addr); | |
1866 | set_mem_attributes (args[i].stack_slot, | |
1867 | TREE_TYPE (args[i].tree_value), 1); | |
1868 | } | |
1869 | set_mem_align (args[i].stack_slot, args[i].locate.boundary); | |
1870 | ||
1871 | /* Function incoming arguments may overlap with sibling call | |
1872 | outgoing arguments and we cannot allow reordering of reads | |
1873 | from function arguments with stores to outgoing arguments | |
1874 | of sibling calls. */ | |
1875 | set_mem_alias_set (args[i].stack, 0); | |
1876 | set_mem_alias_set (args[i].stack_slot, 0); | |
1877 | } | |
1878 | } | |
1879 | } | |
1880 | ||
1881 | /* Given a FNDECL and EXP, return an rtx suitable for use as a target address | |
1882 | in a call instruction. | |
1883 | ||
1884 | FNDECL is the tree node for the target function. For an indirect call | |
1885 | FNDECL will be NULL_TREE. | |
1886 | ||
1887 | ADDR is the operand 0 of CALL_EXPR for this call. */ | |
1888 | ||
1889 | static rtx | |
1890 | rtx_for_function_call (tree fndecl, tree addr) | |
1891 | { | |
1892 | rtx funexp; | |
1893 | ||
1894 | /* Get the function to call, in the form of RTL. */ | |
1895 | if (fndecl) | |
1896 | { | |
1897 | if (!TREE_USED (fndecl) && fndecl != current_function_decl) | |
1898 | TREE_USED (fndecl) = 1; | |
1899 | ||
1900 | /* Get a SYMBOL_REF rtx for the function address. */ | |
1901 | funexp = XEXP (DECL_RTL (fndecl), 0); | |
1902 | } | |
1903 | else | |
1904 | /* Generate an rtx (probably a pseudo-register) for the address. */ | |
1905 | { | |
1906 | push_temp_slots (); | |
1907 | funexp = expand_normal (addr); | |
1908 | pop_temp_slots (); /* FUNEXP can't be BLKmode. */ | |
1909 | } | |
1910 | return funexp; | |
1911 | } | |
1912 | ||
1913 | /* Return the static chain for this function, if any. */ | |
1914 | ||
1915 | rtx | |
1916 | rtx_for_static_chain (const_tree fndecl_or_type, bool incoming_p) | |
1917 | { | |
1918 | if (DECL_P (fndecl_or_type) && !DECL_STATIC_CHAIN (fndecl_or_type)) | |
1919 | return NULL; | |
1920 | ||
1921 | return targetm.calls.static_chain (fndecl_or_type, incoming_p); | |
1922 | } | |
1923 | ||
1924 | /* Internal state for internal_arg_pointer_based_exp and its helpers. */ | |
1925 | static struct | |
1926 | { | |
1927 | /* Last insn that has been scanned by internal_arg_pointer_based_exp_scan, | |
1928 | or NULL_RTX if none has been scanned yet. */ | |
1929 | rtx_insn *scan_start; | |
1930 | /* Vector indexed by REGNO - FIRST_PSEUDO_REGISTER, recording if a pseudo is | |
1931 | based on crtl->args.internal_arg_pointer. The element is NULL_RTX if the | |
1932 | pseudo isn't based on it, a CONST_INT offset if the pseudo is based on it | |
1933 | with fixed offset, or PC if this is with variable or unknown offset. */ | |
1934 | vec<rtx> cache; | |
1935 | } internal_arg_pointer_exp_state; | |
1936 | ||
1937 | static rtx internal_arg_pointer_based_exp (const_rtx, bool); | |
1938 | ||
1939 | /* Helper function for internal_arg_pointer_based_exp. Scan insns in | |
1940 | the tail call sequence, starting with first insn that hasn't been | |
1941 | scanned yet, and note for each pseudo on the LHS whether it is based | |
1942 | on crtl->args.internal_arg_pointer or not, and what offset from that | |
1943 | that pointer it has. */ | |
1944 | ||
1945 | static void | |
1946 | internal_arg_pointer_based_exp_scan (void) | |
1947 | { | |
1948 | rtx_insn *insn, *scan_start = internal_arg_pointer_exp_state.scan_start; | |
1949 | ||
1950 | if (scan_start == NULL_RTX) | |
1951 | insn = get_insns (); | |
1952 | else | |
1953 | insn = NEXT_INSN (scan_start); | |
1954 | ||
1955 | while (insn) | |
1956 | { | |
1957 | rtx set = single_set (insn); | |
1958 | if (set && REG_P (SET_DEST (set)) && !HARD_REGISTER_P (SET_DEST (set))) | |
1959 | { | |
1960 | rtx val = NULL_RTX; | |
1961 | unsigned int idx = REGNO (SET_DEST (set)) - FIRST_PSEUDO_REGISTER; | |
1962 | /* Punt on pseudos set multiple times. */ | |
1963 | if (idx < internal_arg_pointer_exp_state.cache.length () | |
1964 | && (internal_arg_pointer_exp_state.cache[idx] | |
1965 | != NULL_RTX)) | |
1966 | val = pc_rtx; | |
1967 | else | |
1968 | val = internal_arg_pointer_based_exp (SET_SRC (set), false); | |
1969 | if (val != NULL_RTX) | |
1970 | { | |
1971 | if (idx >= internal_arg_pointer_exp_state.cache.length ()) | |
1972 | internal_arg_pointer_exp_state.cache | |
1973 | .safe_grow_cleared (idx + 1, true); | |
1974 | internal_arg_pointer_exp_state.cache[idx] = val; | |
1975 | } | |
1976 | } | |
1977 | if (NEXT_INSN (insn) == NULL_RTX) | |
1978 | scan_start = insn; | |
1979 | insn = NEXT_INSN (insn); | |
1980 | } | |
1981 | ||
1982 | internal_arg_pointer_exp_state.scan_start = scan_start; | |
1983 | } | |
1984 | ||
1985 | /* Compute whether RTL is based on crtl->args.internal_arg_pointer. Return | |
1986 | NULL_RTX if RTL isn't based on it, a CONST_INT offset if RTL is based on | |
1987 | it with fixed offset, or PC if this is with variable or unknown offset. | |
1988 | TOPLEVEL is true if the function is invoked at the topmost level. */ | |
1989 | ||
1990 | static rtx | |
1991 | internal_arg_pointer_based_exp (const_rtx rtl, bool toplevel) | |
1992 | { | |
1993 | if (CONSTANT_P (rtl)) | |
1994 | return NULL_RTX; | |
1995 | ||
1996 | if (rtl == crtl->args.internal_arg_pointer) | |
1997 | return const0_rtx; | |
1998 | ||
1999 | if (REG_P (rtl) && HARD_REGISTER_P (rtl)) | |
2000 | return NULL_RTX; | |
2001 | ||
2002 | poly_int64 offset; | |
2003 | if (GET_CODE (rtl) == PLUS && poly_int_rtx_p (XEXP (rtl, 1), &offset)) | |
2004 | { | |
2005 | rtx val = internal_arg_pointer_based_exp (XEXP (rtl, 0), toplevel); | |
2006 | if (val == NULL_RTX || val == pc_rtx) | |
2007 | return val; | |
2008 | return plus_constant (Pmode, val, offset); | |
2009 | } | |
2010 | ||
2011 | /* When called at the topmost level, scan pseudo assignments in between the | |
2012 | last scanned instruction in the tail call sequence and the latest insn | |
2013 | in that sequence. */ | |
2014 | if (toplevel) | |
2015 | internal_arg_pointer_based_exp_scan (); | |
2016 | ||
2017 | if (REG_P (rtl)) | |
2018 | { | |
2019 | unsigned int idx = REGNO (rtl) - FIRST_PSEUDO_REGISTER; | |
2020 | if (idx < internal_arg_pointer_exp_state.cache.length ()) | |
2021 | return internal_arg_pointer_exp_state.cache[idx]; | |
2022 | ||
2023 | return NULL_RTX; | |
2024 | } | |
2025 | ||
2026 | subrtx_iterator::array_type array; | |
2027 | FOR_EACH_SUBRTX (iter, array, rtl, NONCONST) | |
2028 | { | |
2029 | const_rtx x = *iter; | |
2030 | if (REG_P (x) && internal_arg_pointer_based_exp (x, false) != NULL_RTX) | |
2031 | return pc_rtx; | |
2032 | if (MEM_P (x)) | |
2033 | iter.skip_subrtxes (); | |
2034 | } | |
2035 | ||
2036 | return NULL_RTX; | |
2037 | } | |
2038 | ||
2039 | /* Return true if SIZE bytes starting from address ADDR might overlap an | |
2040 | already-clobbered argument area. This function is used to determine | |
2041 | if we should give up a sibcall. */ | |
2042 | ||
2043 | static bool | |
2044 | mem_might_overlap_already_clobbered_arg_p (rtx addr, poly_uint64 size) | |
2045 | { | |
2046 | poly_int64 i; | |
2047 | unsigned HOST_WIDE_INT start, end; | |
2048 | rtx val; | |
2049 | ||
2050 | if (bitmap_empty_p (stored_args_map) | |
2051 | && stored_args_watermark == HOST_WIDE_INT_M1U) | |
2052 | return false; | |
2053 | val = internal_arg_pointer_based_exp (addr, true); | |
2054 | if (val == NULL_RTX) | |
2055 | return false; | |
2056 | else if (!poly_int_rtx_p (val, &i)) | |
2057 | return true; | |
2058 | ||
2059 | if (known_eq (size, 0U)) | |
2060 | return false; | |
2061 | ||
2062 | if (STACK_GROWS_DOWNWARD) | |
2063 | i -= crtl->args.pretend_args_size; | |
2064 | else | |
2065 | i += crtl->args.pretend_args_size; | |
2066 | ||
2067 | if (ARGS_GROW_DOWNWARD) | |
2068 | i = -i - size; | |
2069 | ||
2070 | /* We can ignore any references to the function's pretend args, | |
2071 | which at this point would manifest as negative values of I. */ | |
2072 | if (known_le (i, 0) && known_le (size, poly_uint64 (-i))) | |
2073 | return false; | |
2074 | ||
2075 | start = maybe_lt (i, 0) ? 0 : constant_lower_bound (i); | |
2076 | if (!(i + size).is_constant (&end)) | |
2077 | end = HOST_WIDE_INT_M1U; | |
2078 | ||
2079 | if (end > stored_args_watermark) | |
2080 | return true; | |
2081 | ||
2082 | end = MIN (end, SBITMAP_SIZE (stored_args_map)); | |
2083 | for (unsigned HOST_WIDE_INT k = start; k < end; ++k) | |
2084 | if (bitmap_bit_p (stored_args_map, k)) | |
2085 | return true; | |
2086 | ||
2087 | return false; | |
2088 | } | |
2089 | ||
2090 | /* Do the register loads required for any wholly-register parms or any | |
2091 | parms which are passed both on the stack and in a register. Their | |
2092 | expressions were already evaluated. | |
2093 | ||
2094 | Mark all register-parms as living through the call, putting these USE | |
2095 | insns in the CALL_INSN_FUNCTION_USAGE field. | |
2096 | ||
2097 | When IS_SIBCALL, perform the check_sibcall_argument_overlap | |
2098 | checking, setting *SIBCALL_FAILURE if appropriate. */ | |
2099 | ||
2100 | static void | |
2101 | load_register_parameters (struct arg_data *args, int num_actuals, | |
2102 | rtx *call_fusage, int flags, int is_sibcall, | |
2103 | bool *sibcall_failure) | |
2104 | { | |
2105 | int i, j; | |
2106 | ||
2107 | for (i = 0; i < num_actuals; i++) | |
2108 | { | |
2109 | rtx reg = ((flags & ECF_SIBCALL) | |
2110 | ? args[i].tail_call_reg : args[i].reg); | |
2111 | if (reg) | |
2112 | { | |
2113 | int partial = args[i].partial; | |
2114 | int nregs; | |
2115 | poly_int64 size = 0; | |
2116 | HOST_WIDE_INT const_size = 0; | |
2117 | rtx_insn *before_arg = get_last_insn (); | |
2118 | tree tree_value = args[i].tree_value; | |
2119 | tree type = TREE_TYPE (tree_value); | |
2120 | if (RECORD_OR_UNION_TYPE_P (type) && TYPE_TRANSPARENT_AGGR (type)) | |
2121 | type = TREE_TYPE (first_field (type)); | |
2122 | /* Set non-negative if we must move a word at a time, even if | |
2123 | just one word (e.g, partial == 4 && mode == DFmode). Set | |
2124 | to -1 if we just use a normal move insn. This value can be | |
2125 | zero if the argument is a zero size structure. */ | |
2126 | nregs = -1; | |
2127 | if (GET_CODE (reg) == PARALLEL) | |
2128 | ; | |
2129 | else if (partial) | |
2130 | { | |
2131 | gcc_assert (partial % UNITS_PER_WORD == 0); | |
2132 | nregs = partial / UNITS_PER_WORD; | |
2133 | } | |
2134 | else if (TYPE_MODE (type) == BLKmode) | |
2135 | { | |
2136 | /* Variable-sized parameters should be described by a | |
2137 | PARALLEL instead. */ | |
2138 | const_size = int_size_in_bytes (type); | |
2139 | gcc_assert (const_size >= 0); | |
2140 | nregs = (const_size + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD; | |
2141 | size = const_size; | |
2142 | } | |
2143 | else | |
2144 | size = GET_MODE_SIZE (args[i].mode); | |
2145 | ||
2146 | /* Handle calls that pass values in multiple non-contiguous | |
2147 | locations. The Irix 6 ABI has examples of this. */ | |
2148 | ||
2149 | if (GET_CODE (reg) == PARALLEL) | |
2150 | emit_group_move (reg, args[i].parallel_value); | |
2151 | ||
2152 | /* If simple case, just do move. If normal partial, store_one_arg | |
2153 | has already loaded the register for us. In all other cases, | |
2154 | load the register(s) from memory. */ | |
2155 | ||
2156 | else if (nregs == -1) | |
2157 | { | |
2158 | emit_move_insn (reg, args[i].value); | |
2159 | #ifdef BLOCK_REG_PADDING | |
2160 | /* Handle case where we have a value that needs shifting | |
2161 | up to the msb. eg. a QImode value and we're padding | |
2162 | upward on a BYTES_BIG_ENDIAN machine. */ | |
2163 | if (args[i].locate.where_pad | |
2164 | == (BYTES_BIG_ENDIAN ? PAD_UPWARD : PAD_DOWNWARD)) | |
2165 | { | |
2166 | gcc_checking_assert (ordered_p (size, UNITS_PER_WORD)); | |
2167 | if (maybe_lt (size, UNITS_PER_WORD)) | |
2168 | { | |
2169 | rtx x; | |
2170 | poly_int64 shift | |
2171 | = (UNITS_PER_WORD - size) * BITS_PER_UNIT; | |
2172 | ||
2173 | /* Assigning REG here rather than a temp makes | |
2174 | CALL_FUSAGE report the whole reg as used. | |
2175 | Strictly speaking, the call only uses SIZE | |
2176 | bytes at the msb end, but it doesn't seem worth | |
2177 | generating rtl to say that. */ | |
2178 | reg = gen_rtx_REG (word_mode, REGNO (reg)); | |
2179 | x = expand_shift (LSHIFT_EXPR, word_mode, | |
2180 | reg, shift, reg, 1); | |
2181 | if (x != reg) | |
2182 | emit_move_insn (reg, x); | |
2183 | } | |
2184 | } | |
2185 | #endif | |
2186 | } | |
2187 | ||
2188 | /* If we have pre-computed the values to put in the registers in | |
2189 | the case of non-aligned structures, copy them in now. */ | |
2190 | ||
2191 | else if (args[i].n_aligned_regs != 0) | |
2192 | for (j = 0; j < args[i].n_aligned_regs; j++) | |
2193 | emit_move_insn (gen_rtx_REG (word_mode, REGNO (reg) + j), | |
2194 | args[i].aligned_regs[j]); | |
2195 | ||
2196 | /* If we need a single register and the source is a constant | |
2197 | VAR_DECL with a simple constructor, expand that constructor | |
2198 | via a pseudo rather than read from (possibly misaligned) | |
2199 | memory. PR middle-end/95126. */ | |
2200 | else if (nregs == 1 | |
2201 | && partial == 0 | |
2202 | && !args[i].pass_on_stack | |
2203 | && VAR_P (tree_value) | |
2204 | && TREE_READONLY (tree_value) | |
2205 | && !TREE_SIDE_EFFECTS (tree_value) | |
2206 | && immediate_const_ctor_p (DECL_INITIAL (tree_value))) | |
2207 | { | |
2208 | rtx target = gen_reg_rtx (word_mode); | |
2209 | store_constructor (DECL_INITIAL (tree_value), target, 0, | |
2210 | int_expr_size (DECL_INITIAL (tree_value)), | |
2211 | false); | |
2212 | reg = gen_rtx_REG (word_mode, REGNO (reg)); | |
2213 | emit_move_insn (reg, target); | |
2214 | } | |
2215 | else if (partial == 0 || args[i].pass_on_stack) | |
2216 | { | |
2217 | /* SIZE and CONST_SIZE are 0 for partial arguments and | |
2218 | the size of a BLKmode type otherwise. */ | |
2219 | gcc_checking_assert (known_eq (size, const_size)); | |
2220 | rtx mem = validize_mem (copy_rtx (args[i].value)); | |
2221 | ||
2222 | /* Check for overlap with already clobbered argument area, | |
2223 | providing that this has non-zero size. */ | |
2224 | if (is_sibcall | |
2225 | && const_size != 0 | |
2226 | && (mem_might_overlap_already_clobbered_arg_p | |
2227 | (XEXP (args[i].value, 0), const_size))) | |
2228 | *sibcall_failure = true; | |
2229 | ||
2230 | if (const_size % UNITS_PER_WORD == 0 | |
2231 | || MEM_ALIGN (mem) % BITS_PER_WORD == 0) | |
2232 | move_block_to_reg (REGNO (reg), mem, nregs, args[i].mode); | |
2233 | else | |
2234 | { | |
2235 | if (nregs > 1) | |
2236 | move_block_to_reg (REGNO (reg), mem, nregs - 1, | |
2237 | args[i].mode); | |
2238 | rtx dest = gen_rtx_REG (word_mode, REGNO (reg) + nregs - 1); | |
2239 | unsigned int bitoff = (nregs - 1) * BITS_PER_WORD; | |
2240 | unsigned int bitsize = const_size * BITS_PER_UNIT - bitoff; | |
2241 | rtx x = extract_bit_field (mem, bitsize, bitoff, 1, dest, | |
2242 | word_mode, word_mode, false, | |
2243 | NULL); | |
2244 | if (BYTES_BIG_ENDIAN) | |
2245 | x = expand_shift (LSHIFT_EXPR, word_mode, x, | |
2246 | BITS_PER_WORD - bitsize, dest, 1); | |
2247 | if (x != dest) | |
2248 | emit_move_insn (dest, x); | |
2249 | } | |
2250 | ||
2251 | /* Handle a BLKmode that needs shifting. */ | |
2252 | if (nregs == 1 && const_size < UNITS_PER_WORD | |
2253 | #ifdef BLOCK_REG_PADDING | |
2254 | && args[i].locate.where_pad == PAD_DOWNWARD | |
2255 | #else | |
2256 | && BYTES_BIG_ENDIAN | |
2257 | #endif | |
2258 | ) | |
2259 | { | |
2260 | rtx dest = gen_rtx_REG (word_mode, REGNO (reg)); | |
2261 | int shift = (UNITS_PER_WORD - const_size) * BITS_PER_UNIT; | |
2262 | enum tree_code dir = (BYTES_BIG_ENDIAN | |
2263 | ? RSHIFT_EXPR : LSHIFT_EXPR); | |
2264 | rtx x; | |
2265 | ||
2266 | x = expand_shift (dir, word_mode, dest, shift, dest, 1); | |
2267 | if (x != dest) | |
2268 | emit_move_insn (dest, x); | |
2269 | } | |
2270 | } | |
2271 | ||
2272 | /* When a parameter is a block, and perhaps in other cases, it is | |
2273 | possible that it did a load from an argument slot that was | |
2274 | already clobbered. */ | |
2275 | if (is_sibcall | |
2276 | && check_sibcall_argument_overlap (before_arg, &args[i], false)) | |
2277 | *sibcall_failure = true; | |
2278 | ||
2279 | /* Handle calls that pass values in multiple non-contiguous | |
2280 | locations. The Irix 6 ABI has examples of this. */ | |
2281 | if (GET_CODE (reg) == PARALLEL) | |
2282 | use_group_regs (call_fusage, reg); | |
2283 | else if (nregs == -1) | |
2284 | use_reg_mode (call_fusage, reg, TYPE_MODE (type)); | |
2285 | else if (nregs > 0) | |
2286 | use_regs (call_fusage, REGNO (reg), nregs); | |
2287 | } | |
2288 | } | |
2289 | } | |
2290 | ||
2291 | /* We need to pop PENDING_STACK_ADJUST bytes. But, if the arguments | |
2292 | wouldn't fill up an even multiple of PREFERRED_UNIT_STACK_BOUNDARY | |
2293 | bytes, then we would need to push some additional bytes to pad the | |
2294 | arguments. So, we try to compute an adjust to the stack pointer for an | |
2295 | amount that will leave the stack under-aligned by UNADJUSTED_ARGS_SIZE | |
2296 | bytes. Then, when the arguments are pushed the stack will be perfectly | |
2297 | aligned. | |
2298 | ||
2299 | Return true if this optimization is possible, storing the adjustment | |
2300 | in ADJUSTMENT_OUT and setting ARGS_SIZE->CONSTANT to the number of | |
2301 | bytes that should be popped after the call. */ | |
2302 | ||
2303 | static bool | |
2304 | combine_pending_stack_adjustment_and_call (poly_int64 *adjustment_out, | |
2305 | poly_int64 unadjusted_args_size, | |
2306 | struct args_size *args_size, | |
2307 | unsigned int preferred_unit_stack_boundary) | |
2308 | { | |
2309 | /* The number of bytes to pop so that the stack will be | |
2310 | under-aligned by UNADJUSTED_ARGS_SIZE bytes. */ | |
2311 | poly_int64 adjustment; | |
2312 | /* The alignment of the stack after the arguments are pushed, if we | |
2313 | just pushed the arguments without adjust the stack here. */ | |
2314 | unsigned HOST_WIDE_INT unadjusted_alignment; | |
2315 | ||
2316 | if (!known_misalignment (stack_pointer_delta + unadjusted_args_size, | |
2317 | preferred_unit_stack_boundary, | |
2318 | &unadjusted_alignment)) | |
2319 | return false; | |
2320 | ||
2321 | /* We want to get rid of as many of the PENDING_STACK_ADJUST bytes | |
2322 | as possible -- leaving just enough left to cancel out the | |
2323 | UNADJUSTED_ALIGNMENT. In other words, we want to ensure that the | |
2324 | PENDING_STACK_ADJUST is non-negative, and congruent to | |
2325 | -UNADJUSTED_ALIGNMENT modulo the PREFERRED_UNIT_STACK_BOUNDARY. */ | |
2326 | ||
2327 | /* Begin by trying to pop all the bytes. */ | |
2328 | unsigned HOST_WIDE_INT tmp_misalignment; | |
2329 | if (!known_misalignment (pending_stack_adjust, | |
2330 | preferred_unit_stack_boundary, | |
2331 | &tmp_misalignment)) | |
2332 | return false; | |
2333 | unadjusted_alignment -= tmp_misalignment; | |
2334 | adjustment = pending_stack_adjust; | |
2335 | /* Push enough additional bytes that the stack will be aligned | |
2336 | after the arguments are pushed. */ | |
2337 | if (preferred_unit_stack_boundary > 1 && unadjusted_alignment) | |
2338 | adjustment -= preferred_unit_stack_boundary - unadjusted_alignment; | |
2339 | ||
2340 | /* We need to know whether the adjusted argument size | |
2341 | (UNADJUSTED_ARGS_SIZE - ADJUSTMENT) constitutes an allocation | |
2342 | or a deallocation. */ | |
2343 | if (!ordered_p (adjustment, unadjusted_args_size)) | |
2344 | return false; | |
2345 | ||
2346 | /* Now, sets ARGS_SIZE->CONSTANT so that we pop the right number of | |
2347 | bytes after the call. The right number is the entire | |
2348 | PENDING_STACK_ADJUST less our ADJUSTMENT plus the amount required | |
2349 | by the arguments in the first place. */ | |
2350 | args_size->constant | |
2351 | = pending_stack_adjust - adjustment + unadjusted_args_size; | |
2352 | ||
2353 | *adjustment_out = adjustment; | |
2354 | return true; | |
2355 | } | |
2356 | ||
2357 | /* Scan X expression if it does not dereference any argument slots | |
2358 | we already clobbered by tail call arguments (as noted in stored_args_map | |
2359 | bitmap). | |
2360 | Return true if X expression dereferences such argument slots, | |
2361 | false otherwise. */ | |
2362 | ||
2363 | static bool | |
2364 | check_sibcall_argument_overlap_1 (rtx x) | |
2365 | { | |
2366 | RTX_CODE code; | |
2367 | int i, j; | |
2368 | const char *fmt; | |
2369 | ||
2370 | if (x == NULL_RTX) | |
2371 | return false; | |
2372 | ||
2373 | code = GET_CODE (x); | |
2374 | ||
2375 | /* We need not check the operands of the CALL expression itself. */ | |
2376 | if (code == CALL) | |
2377 | return false; | |
2378 | ||
2379 | if (code == MEM) | |
2380 | return (mem_might_overlap_already_clobbered_arg_p | |
2381 | (XEXP (x, 0), GET_MODE_SIZE (GET_MODE (x)))); | |
2382 | ||
2383 | /* Scan all subexpressions. */ | |
2384 | fmt = GET_RTX_FORMAT (code); | |
2385 | for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++) | |
2386 | { | |
2387 | if (*fmt == 'e') | |
2388 | { | |
2389 | if (check_sibcall_argument_overlap_1 (XEXP (x, i))) | |
2390 | return true; | |
2391 | } | |
2392 | else if (*fmt == 'E') | |
2393 | { | |
2394 | for (j = 0; j < XVECLEN (x, i); j++) | |
2395 | if (check_sibcall_argument_overlap_1 (XVECEXP (x, i, j))) | |
2396 | return true; | |
2397 | } | |
2398 | } | |
2399 | return false; | |
2400 | } | |
2401 | ||
2402 | /* Scan sequence after INSN if it does not dereference any argument slots | |
2403 | we already clobbered by tail call arguments (as noted in stored_args_map | |
2404 | bitmap). If MARK_STORED_ARGS_MAP, add stack slots for ARG to | |
2405 | stored_args_map bitmap afterwards (when ARG is a register | |
2406 | MARK_STORED_ARGS_MAP should be false). Return true if sequence after | |
2407 | INSN dereferences such argument slots, false otherwise. */ | |
2408 | ||
2409 | static bool | |
2410 | check_sibcall_argument_overlap (rtx_insn *insn, struct arg_data *arg, | |
2411 | bool mark_stored_args_map) | |
2412 | { | |
2413 | poly_uint64 low, high; | |
2414 | unsigned HOST_WIDE_INT const_low, const_high; | |
2415 | ||
2416 | if (insn == NULL_RTX) | |
2417 | insn = get_insns (); | |
2418 | else | |
2419 | insn = NEXT_INSN (insn); | |
2420 | ||
2421 | for (; insn; insn = NEXT_INSN (insn)) | |
2422 | if (INSN_P (insn) | |
2423 | && check_sibcall_argument_overlap_1 (PATTERN (insn))) | |
2424 | break; | |
2425 | ||
2426 | if (mark_stored_args_map) | |
2427 | { | |
2428 | if (ARGS_GROW_DOWNWARD) | |
2429 | low = -arg->locate.slot_offset.constant - arg->locate.size.constant; | |
2430 | else | |
2431 | low = arg->locate.slot_offset.constant; | |
2432 | high = low + arg->locate.size.constant; | |
2433 | ||
2434 | const_low = constant_lower_bound (low); | |
2435 | if (high.is_constant (&const_high)) | |
2436 | for (unsigned HOST_WIDE_INT i = const_low; i < const_high; ++i) | |
2437 | bitmap_set_bit (stored_args_map, i); | |
2438 | else | |
2439 | stored_args_watermark = MIN (stored_args_watermark, const_low); | |
2440 | } | |
2441 | return insn != NULL_RTX; | |
2442 | } | |
2443 | ||
2444 | /* Given that a function returns a value of mode MODE at the most | |
2445 | significant end of hard register VALUE, shift VALUE left or right | |
2446 | as specified by LEFT_P. Return true if some action was needed. */ | |
2447 | ||
2448 | bool | |
2449 | shift_return_value (machine_mode mode, bool left_p, rtx value) | |
2450 | { | |
2451 | gcc_assert (REG_P (value) && HARD_REGISTER_P (value)); | |
2452 | machine_mode value_mode = GET_MODE (value); | |
2453 | poly_int64 shift = GET_MODE_BITSIZE (value_mode) - GET_MODE_BITSIZE (mode); | |
2454 | ||
2455 | if (known_eq (shift, 0)) | |
2456 | return false; | |
2457 | ||
2458 | /* Use ashr rather than lshr for right shifts. This is for the benefit | |
2459 | of the MIPS port, which requires SImode values to be sign-extended | |
2460 | when stored in 64-bit registers. */ | |
2461 | if (!force_expand_binop (value_mode, left_p ? ashl_optab : ashr_optab, | |
2462 | value, gen_int_shift_amount (value_mode, shift), | |
2463 | value, 1, OPTAB_WIDEN)) | |
2464 | gcc_unreachable (); | |
2465 | return true; | |
2466 | } | |
2467 | ||
2468 | /* If X is a likely-spilled register value, copy it to a pseudo | |
2469 | register and return that register. Return X otherwise. */ | |
2470 | ||
2471 | static rtx | |
2472 | avoid_likely_spilled_reg (rtx x) | |
2473 | { | |
2474 | rtx new_rtx; | |
2475 | ||
2476 | if (REG_P (x) | |
2477 | && HARD_REGISTER_P (x) | |
2478 | && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (x)))) | |
2479 | { | |
2480 | /* Make sure that we generate a REG rather than a CONCAT. | |
2481 | Moves into CONCATs can need nontrivial instructions, | |
2482 | and the whole point of this function is to avoid | |
2483 | using the hard register directly in such a situation. */ | |
2484 | generating_concat_p = 0; | |
2485 | new_rtx = gen_reg_rtx (GET_MODE (x)); | |
2486 | generating_concat_p = 1; | |
2487 | emit_move_insn (new_rtx, x); | |
2488 | return new_rtx; | |
2489 | } | |
2490 | return x; | |
2491 | } | |
2492 | ||
2493 | /* Helper function for expand_call. | |
2494 | Return false is EXP is not implementable as a sibling call. */ | |
2495 | ||
2496 | static bool | |
2497 | can_implement_as_sibling_call_p (tree exp, | |
2498 | rtx structure_value_addr, | |
2499 | tree funtype, | |
2500 | tree fndecl, | |
2501 | int flags, | |
2502 | tree addr, | |
2503 | const args_size &args_size) | |
2504 | { | |
2505 | if (!targetm.have_sibcall_epilogue () | |
2506 | && !targetm.emit_epilogue_for_sibcall) | |
2507 | { | |
2508 | maybe_complain_about_tail_call | |
2509 | (exp, | |
2510 | "machine description does not have" | |
2511 | " a sibcall_epilogue instruction pattern"); | |
2512 | return false; | |
2513 | } | |
2514 | ||
2515 | /* Doing sibling call optimization needs some work, since | |
2516 | structure_value_addr can be allocated on the stack. | |
2517 | It does not seem worth the effort since few optimizable | |
2518 | sibling calls will return a structure. */ | |
2519 | if (structure_value_addr != NULL_RTX) | |
2520 | { | |
2521 | maybe_complain_about_tail_call (exp, "callee returns a structure"); | |
2522 | return false; | |
2523 | } | |
2524 | ||
2525 | /* Check whether the target is able to optimize the call | |
2526 | into a sibcall. */ | |
2527 | if (!targetm.function_ok_for_sibcall (fndecl, exp)) | |
2528 | { | |
2529 | maybe_complain_about_tail_call (exp, | |
2530 | "target is not able to optimize the" | |
2531 | " call into a sibling call"); | |
2532 | return false; | |
2533 | } | |
2534 | ||
2535 | /* Functions that do not return exactly once may not be sibcall | |
2536 | optimized. */ | |
2537 | if (flags & ECF_RETURNS_TWICE) | |
2538 | { | |
2539 | maybe_complain_about_tail_call (exp, "callee returns twice"); | |
2540 | return false; | |
2541 | } | |
2542 | if (flags & ECF_NORETURN) | |
2543 | { | |
2544 | maybe_complain_about_tail_call (exp, "callee does not return"); | |
2545 | return false; | |
2546 | } | |
2547 | ||
2548 | if (TYPE_VOLATILE (TREE_TYPE (TREE_TYPE (addr)))) | |
2549 | { | |
2550 | maybe_complain_about_tail_call (exp, "volatile function type"); | |
2551 | return false; | |
2552 | } | |
2553 | ||
2554 | /* __sanitizer_cov_trace_pc is supposed to inspect its return address | |
2555 | to identify the caller, and therefore should not be tailcalled. */ | |
2556 | if (fndecl && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL | |
2557 | && DECL_FUNCTION_CODE (fndecl) == BUILT_IN_SANITIZER_COV_TRACE_PC) | |
2558 | { | |
2559 | /* No need for maybe_complain_about_tail_call here: | |
2560 | the call is synthesized by the compiler. */ | |
2561 | return false; | |
2562 | } | |
2563 | ||
2564 | /* If the called function is nested in the current one, it might access | |
2565 | some of the caller's arguments, but could clobber them beforehand if | |
2566 | the argument areas are shared. */ | |
2567 | if (fndecl && decl_function_context (fndecl) == current_function_decl) | |
2568 | { | |
2569 | maybe_complain_about_tail_call (exp, "nested function"); | |
2570 | return false; | |
2571 | } | |
2572 | ||
2573 | /* If this function requires more stack slots than the current | |
2574 | function, we cannot change it into a sibling call. | |
2575 | crtl->args.pretend_args_size is not part of the | |
2576 | stack allocated by our caller. */ | |
2577 | if (maybe_gt (args_size.constant, | |
2578 | crtl->args.size - crtl->args.pretend_args_size)) | |
2579 | { | |
2580 | maybe_complain_about_tail_call (exp, | |
2581 | "callee required more stack slots" | |
2582 | " than the caller"); | |
2583 | return false; | |
2584 | } | |
2585 | ||
2586 | /* If the callee pops its own arguments, then it must pop exactly | |
2587 | the same number of arguments as the current function. */ | |
2588 | if (maybe_ne (targetm.calls.return_pops_args (fndecl, funtype, | |
2589 | args_size.constant), | |
2590 | targetm.calls.return_pops_args (current_function_decl, | |
2591 | TREE_TYPE | |
2592 | (current_function_decl), | |
2593 | crtl->args.size))) | |
2594 | { | |
2595 | maybe_complain_about_tail_call (exp, | |
2596 | "inconsistent number of" | |
2597 | " popped arguments"); | |
2598 | return false; | |
2599 | } | |
2600 | ||
2601 | if (!lang_hooks.decls.ok_for_sibcall (fndecl)) | |
2602 | { | |
2603 | maybe_complain_about_tail_call (exp, "frontend does not support" | |
2604 | " sibling call"); | |
2605 | return false; | |
2606 | } | |
2607 | ||
2608 | /* All checks passed. */ | |
2609 | return true; | |
2610 | } | |
2611 | ||
2612 | /* Update stack alignment when the parameter is passed in the stack | |
2613 | since the outgoing parameter requires extra alignment on the calling | |
2614 | function side. */ | |
2615 | ||
2616 | static void | |
2617 | update_stack_alignment_for_call (struct locate_and_pad_arg_data *locate) | |
2618 | { | |
2619 | if (crtl->stack_alignment_needed < locate->boundary) | |
2620 | crtl->stack_alignment_needed = locate->boundary; | |
2621 | if (crtl->preferred_stack_boundary < locate->boundary) | |
2622 | crtl->preferred_stack_boundary = locate->boundary; | |
2623 | } | |
2624 | ||
2625 | /* Generate all the code for a CALL_EXPR exp | |
2626 | and return an rtx for its value. | |
2627 | Store the value in TARGET (specified as an rtx) if convenient. | |
2628 | If the value is stored in TARGET then TARGET is returned. | |
2629 | If IGNORE is nonzero, then we ignore the value of the function call. */ | |
2630 | ||
2631 | rtx | |
2632 | expand_call (tree exp, rtx target, int ignore) | |
2633 | { | |
2634 | /* Nonzero if we are currently expanding a call. */ | |
2635 | static int currently_expanding_call = 0; | |
2636 | ||
2637 | /* RTX for the function to be called. */ | |
2638 | rtx funexp; | |
2639 | /* Sequence of insns to perform a normal "call". */ | |
2640 | rtx_insn *normal_call_insns = NULL; | |
2641 | /* Sequence of insns to perform a tail "call". */ | |
2642 | rtx_insn *tail_call_insns = NULL; | |
2643 | /* Data type of the function. */ | |
2644 | tree funtype; | |
2645 | tree type_arg_types; | |
2646 | tree rettype; | |
2647 | /* Declaration of the function being called, | |
2648 | or 0 if the function is computed (not known by name). */ | |
2649 | tree fndecl = 0; | |
2650 | /* The type of the function being called. */ | |
2651 | tree fntype; | |
2652 | bool try_tail_call = CALL_EXPR_TAILCALL (exp); | |
2653 | bool must_tail_call = CALL_EXPR_MUST_TAIL_CALL (exp); | |
2654 | int pass; | |
2655 | ||
2656 | /* Register in which non-BLKmode value will be returned, | |
2657 | or 0 if no value or if value is BLKmode. */ | |
2658 | rtx valreg; | |
2659 | /* Address where we should return a BLKmode value; | |
2660 | 0 if value not BLKmode. */ | |
2661 | rtx structure_value_addr = 0; | |
2662 | /* Nonzero if that address is being passed by treating it as | |
2663 | an extra, implicit first parameter. Otherwise, | |
2664 | it is passed by being copied directly into struct_value_rtx. */ | |
2665 | int structure_value_addr_parm = 0; | |
2666 | /* Holds the value of implicit argument for the struct value. */ | |
2667 | tree structure_value_addr_value = NULL_TREE; | |
2668 | /* Size of aggregate value wanted, or zero if none wanted | |
2669 | or if we are using the non-reentrant PCC calling convention | |
2670 | or expecting the value in registers. */ | |
2671 | poly_int64 struct_value_size = 0; | |
2672 | /* True if called function returns an aggregate in memory PCC style, | |
2673 | by returning the address of where to find it. */ | |
2674 | bool pcc_struct_value = false; | |
2675 | rtx struct_value = 0; | |
2676 | ||
2677 | /* Number of actual parameters in this call, including struct value addr. */ | |
2678 | int num_actuals; | |
2679 | /* Number of named args. Args after this are anonymous ones | |
2680 | and they must all go on the stack. */ | |
2681 | int n_named_args; | |
2682 | /* Number of complex actual arguments that need to be split. */ | |
2683 | int num_complex_actuals = 0; | |
2684 | ||
2685 | /* Vector of information about each argument. | |
2686 | Arguments are numbered in the order they will be pushed, | |
2687 | not the order they are written. */ | |
2688 | struct arg_data *args; | |
2689 | ||
2690 | /* Total size in bytes of all the stack-parms scanned so far. */ | |
2691 | struct args_size args_size; | |
2692 | struct args_size adjusted_args_size; | |
2693 | /* Size of arguments before any adjustments (such as rounding). */ | |
2694 | poly_int64 unadjusted_args_size; | |
2695 | /* Data on reg parms scanned so far. */ | |
2696 | CUMULATIVE_ARGS args_so_far_v; | |
2697 | cumulative_args_t args_so_far; | |
2698 | /* Nonzero if a reg parm has been scanned. */ | |
2699 | int reg_parm_seen; | |
2700 | ||
2701 | /* True if we must avoid push-insns in the args for this call. | |
2702 | If stack space is allocated for register parameters, but not by the | |
2703 | caller, then it is preallocated in the fixed part of the stack frame. | |
2704 | So the entire argument block must then be preallocated (i.e., we | |
2705 | ignore PUSH_ROUNDING in that case). */ | |
2706 | bool must_preallocate = !targetm.calls.push_argument (0); | |
2707 | ||
2708 | /* Size of the stack reserved for parameter registers. */ | |
2709 | int reg_parm_stack_space = 0; | |
2710 | ||
2711 | /* Address of space preallocated for stack parms | |
2712 | (on machines that lack push insns), or 0 if space not preallocated. */ | |
2713 | rtx argblock = 0; | |
2714 | ||
2715 | /* Mask of ECF_ and ERF_ flags. */ | |
2716 | int flags = 0; | |
2717 | int return_flags = 0; | |
2718 | #ifdef REG_PARM_STACK_SPACE | |
2719 | /* Define the boundary of the register parm stack space that needs to be | |
2720 | saved, if any. */ | |
2721 | int low_to_save, high_to_save; | |
2722 | rtx save_area = 0; /* Place that it is saved */ | |
2723 | #endif | |
2724 | ||
2725 | unsigned int initial_highest_arg_in_use = highest_outgoing_arg_in_use; | |
2726 | char *initial_stack_usage_map = stack_usage_map; | |
2727 | unsigned HOST_WIDE_INT initial_stack_usage_watermark = stack_usage_watermark; | |
2728 | char *stack_usage_map_buf = NULL; | |
2729 | ||
2730 | poly_int64 old_stack_allocated; | |
2731 | ||
2732 | /* State variables to track stack modifications. */ | |
2733 | rtx old_stack_level = 0; | |
2734 | int old_stack_arg_under_construction = 0; | |
2735 | poly_int64 old_pending_adj = 0; | |
2736 | int old_inhibit_defer_pop = inhibit_defer_pop; | |
2737 | ||
2738 | /* Some stack pointer alterations we make are performed via | |
2739 | allocate_dynamic_stack_space. This modifies the stack_pointer_delta, | |
2740 | which we then also need to save/restore along the way. */ | |
2741 | poly_int64 old_stack_pointer_delta = 0; | |
2742 | ||
2743 | rtx call_fusage; | |
2744 | tree addr = CALL_EXPR_FN (exp); | |
2745 | int i; | |
2746 | /* The alignment of the stack, in bits. */ | |
2747 | unsigned HOST_WIDE_INT preferred_stack_boundary; | |
2748 | /* The alignment of the stack, in bytes. */ | |
2749 | unsigned HOST_WIDE_INT preferred_unit_stack_boundary; | |
2750 | /* The static chain value to use for this call. */ | |
2751 | rtx static_chain_value; | |
2752 | /* See if this is "nothrow" function call. */ | |
2753 | if (TREE_NOTHROW (exp)) | |
2754 | flags |= ECF_NOTHROW; | |
2755 | ||
2756 | /* See if we can find a DECL-node for the actual function, and get the | |
2757 | function attributes (flags) from the function decl or type node. */ | |
2758 | fndecl = get_callee_fndecl (exp); | |
2759 | if (fndecl) | |
2760 | { | |
2761 | fntype = TREE_TYPE (fndecl); | |
2762 | flags |= flags_from_decl_or_type (fndecl); | |
2763 | return_flags |= decl_return_flags (fndecl); | |
2764 | } | |
2765 | else | |
2766 | { | |
2767 | fntype = TREE_TYPE (TREE_TYPE (addr)); | |
2768 | flags |= flags_from_decl_or_type (fntype); | |
2769 | if (CALL_EXPR_BY_DESCRIPTOR (exp)) | |
2770 | flags |= ECF_BY_DESCRIPTOR; | |
2771 | } | |
2772 | rettype = TREE_TYPE (exp); | |
2773 | ||
2774 | struct_value = targetm.calls.struct_value_rtx (fntype, 0); | |
2775 | ||
2776 | /* Warn if this value is an aggregate type, | |
2777 | regardless of which calling convention we are using for it. */ | |
2778 | if (AGGREGATE_TYPE_P (rettype)) | |
2779 | warning (OPT_Waggregate_return, "function call has aggregate value"); | |
2780 | ||
2781 | /* If the result of a non looping pure or const function call is | |
2782 | ignored (or void), and none of its arguments are volatile, we can | |
2783 | avoid expanding the call and just evaluate the arguments for | |
2784 | side-effects. */ | |
2785 | if ((flags & (ECF_CONST | ECF_PURE)) | |
2786 | && (!(flags & ECF_LOOPING_CONST_OR_PURE)) | |
2787 | && (flags & ECF_NOTHROW) | |
2788 | && (ignore || target == const0_rtx | |
2789 | || TYPE_MODE (rettype) == VOIDmode)) | |
2790 | { | |
2791 | bool volatilep = false; | |
2792 | tree arg; | |
2793 | call_expr_arg_iterator iter; | |
2794 | ||
2795 | FOR_EACH_CALL_EXPR_ARG (arg, iter, exp) | |
2796 | if (TREE_THIS_VOLATILE (arg)) | |
2797 | { | |
2798 | volatilep = true; | |
2799 | break; | |
2800 | } | |
2801 | ||
2802 | if (! volatilep) | |
2803 | { | |
2804 | FOR_EACH_CALL_EXPR_ARG (arg, iter, exp) | |
2805 | expand_expr (arg, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
2806 | return const0_rtx; | |
2807 | } | |
2808 | } | |
2809 | ||
2810 | #ifdef REG_PARM_STACK_SPACE | |
2811 | reg_parm_stack_space = REG_PARM_STACK_SPACE (!fndecl ? fntype : fndecl); | |
2812 | #endif | |
2813 | ||
2814 | if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))) | |
2815 | && reg_parm_stack_space > 0 && targetm.calls.push_argument (0)) | |
2816 | must_preallocate = true; | |
2817 | ||
2818 | /* Set up a place to return a structure. */ | |
2819 | ||
2820 | /* Cater to broken compilers. */ | |
2821 | if (aggregate_value_p (exp, fntype)) | |
2822 | { | |
2823 | /* This call returns a big structure. */ | |
2824 | flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE); | |
2825 | ||
2826 | #ifdef PCC_STATIC_STRUCT_RETURN | |
2827 | { | |
2828 | pcc_struct_value = true; | |
2829 | } | |
2830 | #else /* not PCC_STATIC_STRUCT_RETURN */ | |
2831 | { | |
2832 | if (!poly_int_tree_p (TYPE_SIZE_UNIT (rettype), &struct_value_size)) | |
2833 | struct_value_size = -1; | |
2834 | ||
2835 | /* Even if it is semantically safe to use the target as the return | |
2836 | slot, it may be not sufficiently aligned for the return type. */ | |
2837 | if (CALL_EXPR_RETURN_SLOT_OPT (exp) | |
2838 | && target | |
2839 | && MEM_P (target) | |
2840 | /* If rettype is addressable, we may not create a temporary. | |
2841 | If target is properly aligned at runtime and the compiler | |
2842 | just doesn't know about it, it will work fine, otherwise it | |
2843 | will be UB. */ | |
2844 | && (TREE_ADDRESSABLE (rettype) | |
2845 | || !(MEM_ALIGN (target) < TYPE_ALIGN (rettype) | |
2846 | && targetm.slow_unaligned_access (TYPE_MODE (rettype), | |
2847 | MEM_ALIGN (target))))) | |
2848 | structure_value_addr = XEXP (target, 0); | |
2849 | else | |
2850 | { | |
2851 | /* For variable-sized objects, we must be called with a target | |
2852 | specified. If we were to allocate space on the stack here, | |
2853 | we would have no way of knowing when to free it. */ | |
2854 | rtx d = assign_temp (rettype, 1, 1); | |
2855 | structure_value_addr = XEXP (d, 0); | |
2856 | target = 0; | |
2857 | } | |
2858 | } | |
2859 | #endif /* not PCC_STATIC_STRUCT_RETURN */ | |
2860 | } | |
2861 | ||
2862 | /* Figure out the amount to which the stack should be aligned. */ | |
2863 | preferred_stack_boundary = PREFERRED_STACK_BOUNDARY; | |
2864 | if (fndecl) | |
2865 | { | |
2866 | struct cgraph_rtl_info *i = cgraph_node::rtl_info (fndecl); | |
2867 | /* Without automatic stack alignment, we can't increase preferred | |
2868 | stack boundary. With automatic stack alignment, it is | |
2869 | unnecessary since unless we can guarantee that all callers will | |
2870 | align the outgoing stack properly, callee has to align its | |
2871 | stack anyway. */ | |
2872 | if (i | |
2873 | && i->preferred_incoming_stack_boundary | |
2874 | && i->preferred_incoming_stack_boundary < preferred_stack_boundary) | |
2875 | preferred_stack_boundary = i->preferred_incoming_stack_boundary; | |
2876 | } | |
2877 | ||
2878 | /* Operand 0 is a pointer-to-function; get the type of the function. */ | |
2879 | funtype = TREE_TYPE (addr); | |
2880 | gcc_assert (POINTER_TYPE_P (funtype)); | |
2881 | funtype = TREE_TYPE (funtype); | |
2882 | ||
2883 | /* Count whether there are actual complex arguments that need to be split | |
2884 | into their real and imaginary parts. Munge the type_arg_types | |
2885 | appropriately here as well. */ | |
2886 | if (targetm.calls.split_complex_arg) | |
2887 | { | |
2888 | call_expr_arg_iterator iter; | |
2889 | tree arg; | |
2890 | FOR_EACH_CALL_EXPR_ARG (arg, iter, exp) | |
2891 | { | |
2892 | tree type = TREE_TYPE (arg); | |
2893 | if (type && TREE_CODE (type) == COMPLEX_TYPE | |
2894 | && targetm.calls.split_complex_arg (type)) | |
2895 | num_complex_actuals++; | |
2896 | } | |
2897 | type_arg_types = split_complex_types (TYPE_ARG_TYPES (funtype)); | |
2898 | } | |
2899 | else | |
2900 | type_arg_types = TYPE_ARG_TYPES (funtype); | |
2901 | ||
2902 | if (flags & ECF_MAY_BE_ALLOCA) | |
2903 | cfun->calls_alloca = 1; | |
2904 | ||
2905 | /* If struct_value_rtx is 0, it means pass the address | |
2906 | as if it were an extra parameter. Put the argument expression | |
2907 | in structure_value_addr_value. */ | |
2908 | if (structure_value_addr && struct_value == 0) | |
2909 | { | |
2910 | /* If structure_value_addr is a REG other than | |
2911 | virtual_outgoing_args_rtx, we can use always use it. If it | |
2912 | is not a REG, we must always copy it into a register. | |
2913 | If it is virtual_outgoing_args_rtx, we must copy it to another | |
2914 | register in some cases. */ | |
2915 | rtx temp = (!REG_P (structure_value_addr) | |
2916 | || (ACCUMULATE_OUTGOING_ARGS | |
2917 | && stack_arg_under_construction | |
2918 | && structure_value_addr == virtual_outgoing_args_rtx) | |
2919 | ? copy_addr_to_reg (convert_memory_address | |
2920 | (Pmode, structure_value_addr)) | |
2921 | : structure_value_addr); | |
2922 | ||
2923 | structure_value_addr_value = | |
2924 | make_tree (build_pointer_type (TREE_TYPE (funtype)), temp); | |
2925 | structure_value_addr_parm = 1; | |
2926 | } | |
2927 | ||
2928 | /* Count the arguments and set NUM_ACTUALS. */ | |
2929 | num_actuals | |
2930 | = call_expr_nargs (exp) + num_complex_actuals + structure_value_addr_parm; | |
2931 | ||
2932 | /* Compute number of named args. | |
2933 | First, do a raw count of the args for INIT_CUMULATIVE_ARGS. */ | |
2934 | ||
2935 | if (type_arg_types != 0) | |
2936 | n_named_args | |
2937 | = (list_length (type_arg_types) | |
2938 | /* Count the struct value address, if it is passed as a parm. */ | |
2939 | + structure_value_addr_parm); | |
2940 | else if (TYPE_NO_NAMED_ARGS_STDARG_P (funtype)) | |
2941 | n_named_args = 0; | |
2942 | else | |
2943 | /* If we know nothing, treat all args as named. */ | |
2944 | n_named_args = num_actuals; | |
2945 | ||
2946 | /* Start updating where the next arg would go. | |
2947 | ||
2948 | On some machines (such as the PA) indirect calls have a different | |
2949 | calling convention than normal calls. The fourth argument in | |
2950 | INIT_CUMULATIVE_ARGS tells the backend if this is an indirect call | |
2951 | or not. */ | |
2952 | INIT_CUMULATIVE_ARGS (args_so_far_v, funtype, NULL_RTX, fndecl, n_named_args); | |
2953 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
2954 | ||
2955 | /* Now possibly adjust the number of named args. | |
2956 | Normally, don't include the last named arg if anonymous args follow. | |
2957 | We do include the last named arg if | |
2958 | targetm.calls.strict_argument_naming() returns nonzero. | |
2959 | (If no anonymous args follow, the result of list_length is actually | |
2960 | one too large. This is harmless.) | |
2961 | ||
2962 | If targetm.calls.pretend_outgoing_varargs_named() returns | |
2963 | nonzero, and targetm.calls.strict_argument_naming() returns zero, | |
2964 | this machine will be able to place unnamed args that were passed | |
2965 | in registers into the stack. So treat all args as named. This | |
2966 | allows the insns emitting for a specific argument list to be | |
2967 | independent of the function declaration. | |
2968 | ||
2969 | If targetm.calls.pretend_outgoing_varargs_named() returns zero, | |
2970 | we do not have any reliable way to pass unnamed args in | |
2971 | registers, so we must force them into memory. */ | |
2972 | ||
2973 | if (type_arg_types != 0 | |
2974 | && targetm.calls.strict_argument_naming (args_so_far)) | |
2975 | ; | |
2976 | else if (type_arg_types != 0 | |
2977 | && ! targetm.calls.pretend_outgoing_varargs_named (args_so_far)) | |
2978 | /* Don't include the last named arg. */ | |
2979 | --n_named_args; | |
2980 | else if (TYPE_NO_NAMED_ARGS_STDARG_P (funtype)) | |
2981 | n_named_args = 0; | |
2982 | else | |
2983 | /* Treat all args as named. */ | |
2984 | n_named_args = num_actuals; | |
2985 | ||
2986 | /* Make a vector to hold all the information about each arg. */ | |
2987 | args = XCNEWVEC (struct arg_data, num_actuals); | |
2988 | ||
2989 | /* Build up entries in the ARGS array, compute the size of the | |
2990 | arguments into ARGS_SIZE, etc. */ | |
2991 | initialize_argument_information (num_actuals, args, &args_size, | |
2992 | n_named_args, exp, | |
2993 | structure_value_addr_value, fndecl, fntype, | |
2994 | args_so_far, reg_parm_stack_space, | |
2995 | &old_stack_level, &old_pending_adj, | |
2996 | &must_preallocate, &flags, | |
2997 | &try_tail_call, CALL_FROM_THUNK_P (exp)); | |
2998 | ||
2999 | if (args_size.var) | |
3000 | must_preallocate = true; | |
3001 | ||
3002 | /* Now make final decision about preallocating stack space. */ | |
3003 | must_preallocate = finalize_must_preallocate (must_preallocate, | |
3004 | num_actuals, args, | |
3005 | &args_size); | |
3006 | ||
3007 | /* If the structure value address will reference the stack pointer, we | |
3008 | must stabilize it. We don't need to do this if we know that we are | |
3009 | not going to adjust the stack pointer in processing this call. */ | |
3010 | ||
3011 | if (structure_value_addr | |
3012 | && (reg_mentioned_p (virtual_stack_dynamic_rtx, structure_value_addr) | |
3013 | || reg_mentioned_p (virtual_outgoing_args_rtx, | |
3014 | structure_value_addr)) | |
3015 | && (args_size.var | |
3016 | || (!ACCUMULATE_OUTGOING_ARGS | |
3017 | && maybe_ne (args_size.constant, 0)))) | |
3018 | structure_value_addr = copy_to_reg (structure_value_addr); | |
3019 | ||
3020 | /* Tail calls can make things harder to debug, and we've traditionally | |
3021 | pushed these optimizations into -O2. Don't try if we're already | |
3022 | expanding a call, as that means we're an argument. Don't try if | |
3023 | there's cleanups, as we know there's code to follow the call. */ | |
3024 | if (currently_expanding_call++ != 0 | |
3025 | || (!flag_optimize_sibling_calls && !CALL_FROM_THUNK_P (exp)) | |
3026 | || args_size.var | |
3027 | || dbg_cnt (tail_call) == false) | |
3028 | try_tail_call = 0; | |
3029 | ||
3030 | /* Workaround buggy C/C++ wrappers around Fortran routines with | |
3031 | character(len=constant) arguments if the hidden string length arguments | |
3032 | are passed on the stack; if the callers forget to pass those arguments, | |
3033 | attempting to tail call in such routines leads to stack corruption. | |
3034 | Avoid tail calls in functions where at least one such hidden string | |
3035 | length argument is passed (partially or fully) on the stack in the | |
3036 | caller and the callee needs to pass any arguments on the stack. | |
3037 | See PR90329. */ | |
3038 | if (try_tail_call && maybe_ne (args_size.constant, 0)) | |
3039 | for (tree arg = DECL_ARGUMENTS (current_function_decl); | |
3040 | arg; arg = DECL_CHAIN (arg)) | |
3041 | if (DECL_HIDDEN_STRING_LENGTH (arg) && DECL_INCOMING_RTL (arg)) | |
3042 | { | |
3043 | subrtx_iterator::array_type array; | |
3044 | FOR_EACH_SUBRTX (iter, array, DECL_INCOMING_RTL (arg), NONCONST) | |
3045 | if (MEM_P (*iter)) | |
3046 | { | |
3047 | try_tail_call = 0; | |
3048 | break; | |
3049 | } | |
3050 | } | |
3051 | ||
3052 | /* If the user has marked the function as requiring tail-call | |
3053 | optimization, attempt it. */ | |
3054 | if (must_tail_call) | |
3055 | try_tail_call = 1; | |
3056 | ||
3057 | /* Rest of purposes for tail call optimizations to fail. */ | |
3058 | if (try_tail_call) | |
3059 | try_tail_call = can_implement_as_sibling_call_p (exp, | |
3060 | structure_value_addr, | |
3061 | funtype, | |
3062 | fndecl, | |
3063 | flags, addr, args_size); | |
3064 | ||
3065 | /* Check if caller and callee disagree in promotion of function | |
3066 | return value. */ | |
3067 | if (try_tail_call) | |
3068 | { | |
3069 | machine_mode caller_mode, caller_promoted_mode; | |
3070 | machine_mode callee_mode, callee_promoted_mode; | |
3071 | int caller_unsignedp, callee_unsignedp; | |
3072 | tree caller_res = DECL_RESULT (current_function_decl); | |
3073 | ||
3074 | caller_unsignedp = TYPE_UNSIGNED (TREE_TYPE (caller_res)); | |
3075 | caller_mode = DECL_MODE (caller_res); | |
3076 | callee_unsignedp = TYPE_UNSIGNED (TREE_TYPE (funtype)); | |
3077 | callee_mode = TYPE_MODE (TREE_TYPE (funtype)); | |
3078 | caller_promoted_mode | |
3079 | = promote_function_mode (TREE_TYPE (caller_res), caller_mode, | |
3080 | &caller_unsignedp, | |
3081 | TREE_TYPE (current_function_decl), 1); | |
3082 | callee_promoted_mode | |
3083 | = promote_function_mode (TREE_TYPE (funtype), callee_mode, | |
3084 | &callee_unsignedp, | |
3085 | funtype, 1); | |
3086 | if (caller_mode != VOIDmode | |
3087 | && (caller_promoted_mode != callee_promoted_mode | |
3088 | || ((caller_mode != caller_promoted_mode | |
3089 | || callee_mode != callee_promoted_mode) | |
3090 | && (caller_unsignedp != callee_unsignedp | |
3091 | || partial_subreg_p (caller_mode, callee_mode))))) | |
3092 | { | |
3093 | try_tail_call = 0; | |
3094 | maybe_complain_about_tail_call (exp, | |
3095 | "caller and callee disagree in" | |
3096 | " promotion of function" | |
3097 | " return value"); | |
3098 | } | |
3099 | } | |
3100 | ||
3101 | /* Ensure current function's preferred stack boundary is at least | |
3102 | what we need. Stack alignment may also increase preferred stack | |
3103 | boundary. */ | |
3104 | for (i = 0; i < num_actuals; i++) | |
3105 | if (reg_parm_stack_space > 0 | |
3106 | || args[i].reg == 0 | |
3107 | || args[i].partial != 0 | |
3108 | || args[i].pass_on_stack) | |
3109 | update_stack_alignment_for_call (&args[i].locate); | |
3110 | if (crtl->preferred_stack_boundary < preferred_stack_boundary) | |
3111 | crtl->preferred_stack_boundary = preferred_stack_boundary; | |
3112 | else | |
3113 | preferred_stack_boundary = crtl->preferred_stack_boundary; | |
3114 | ||
3115 | preferred_unit_stack_boundary = preferred_stack_boundary / BITS_PER_UNIT; | |
3116 | ||
3117 | if (flag_callgraph_info) | |
3118 | record_final_call (fndecl, EXPR_LOCATION (exp)); | |
3119 | ||
3120 | /* We want to make two insn chains; one for a sibling call, the other | |
3121 | for a normal call. We will select one of the two chains after | |
3122 | initial RTL generation is complete. */ | |
3123 | for (pass = try_tail_call ? 0 : 1; pass < 2; pass++) | |
3124 | { | |
3125 | bool sibcall_failure = false; | |
3126 | bool normal_failure = false; | |
3127 | /* We want to emit any pending stack adjustments before the tail | |
3128 | recursion "call". That way we know any adjustment after the tail | |
3129 | recursion call can be ignored if we indeed use the tail | |
3130 | call expansion. */ | |
3131 | saved_pending_stack_adjust save; | |
3132 | rtx_insn *insns, *before_call, *after_args; | |
3133 | rtx next_arg_reg; | |
3134 | ||
3135 | if (pass == 0) | |
3136 | { | |
3137 | /* State variables we need to save and restore between | |
3138 | iterations. */ | |
3139 | save_pending_stack_adjust (&save); | |
3140 | } | |
3141 | if (pass) | |
3142 | flags &= ~ECF_SIBCALL; | |
3143 | else | |
3144 | flags |= ECF_SIBCALL; | |
3145 | ||
3146 | /* Other state variables that we must reinitialize each time | |
3147 | through the loop (that are not initialized by the loop itself). */ | |
3148 | argblock = 0; | |
3149 | call_fusage = 0; | |
3150 | ||
3151 | /* Start a new sequence for the normal call case. | |
3152 | ||
3153 | From this point on, if the sibling call fails, we want to set | |
3154 | sibcall_failure instead of continuing the loop. */ | |
3155 | start_sequence (); | |
3156 | ||
3157 | /* Don't let pending stack adjusts add up to too much. | |
3158 | Also, do all pending adjustments now if there is any chance | |
3159 | this might be a call to alloca or if we are expanding a sibling | |
3160 | call sequence. | |
3161 | Also do the adjustments before a throwing call, otherwise | |
3162 | exception handling can fail; PR 19225. */ | |
3163 | if (maybe_ge (pending_stack_adjust, 32) | |
3164 | || (maybe_ne (pending_stack_adjust, 0) | |
3165 | && (flags & ECF_MAY_BE_ALLOCA)) | |
3166 | || (maybe_ne (pending_stack_adjust, 0) | |
3167 | && flag_exceptions && !(flags & ECF_NOTHROW)) | |
3168 | || pass == 0) | |
3169 | do_pending_stack_adjust (); | |
3170 | ||
3171 | /* Precompute any arguments as needed. */ | |
3172 | if (pass) | |
3173 | precompute_arguments (num_actuals, args); | |
3174 | ||
3175 | /* Now we are about to start emitting insns that can be deleted | |
3176 | if a libcall is deleted. */ | |
3177 | if (pass && (flags & ECF_MALLOC)) | |
3178 | start_sequence (); | |
3179 | ||
3180 | /* Check the canary value for sibcall or function which doesn't | |
3181 | return and could throw. */ | |
3182 | if ((pass == 0 | |
3183 | || ((flags & ECF_NORETURN) != 0 && tree_could_throw_p (exp))) | |
3184 | && crtl->stack_protect_guard | |
3185 | && targetm.stack_protect_runtime_enabled_p ()) | |
3186 | stack_protect_epilogue (); | |
3187 | ||
3188 | adjusted_args_size = args_size; | |
3189 | /* Compute the actual size of the argument block required. The variable | |
3190 | and constant sizes must be combined, the size may have to be rounded, | |
3191 | and there may be a minimum required size. When generating a sibcall | |
3192 | pattern, do not round up, since we'll be re-using whatever space our | |
3193 | caller provided. */ | |
3194 | unadjusted_args_size | |
3195 | = compute_argument_block_size (reg_parm_stack_space, | |
3196 | &adjusted_args_size, | |
3197 | fndecl, fntype, | |
3198 | (pass == 0 ? 0 | |
3199 | : preferred_stack_boundary)); | |
3200 | ||
3201 | old_stack_allocated = stack_pointer_delta - pending_stack_adjust; | |
3202 | ||
3203 | /* The argument block when performing a sibling call is the | |
3204 | incoming argument block. */ | |
3205 | if (pass == 0) | |
3206 | { | |
3207 | argblock = crtl->args.internal_arg_pointer; | |
3208 | if (STACK_GROWS_DOWNWARD) | |
3209 | argblock | |
3210 | = plus_constant (Pmode, argblock, crtl->args.pretend_args_size); | |
3211 | else | |
3212 | argblock | |
3213 | = plus_constant (Pmode, argblock, -crtl->args.pretend_args_size); | |
3214 | ||
3215 | HOST_WIDE_INT map_size = constant_lower_bound (args_size.constant); | |
3216 | stored_args_map = sbitmap_alloc (map_size); | |
3217 | bitmap_clear (stored_args_map); | |
3218 | stored_args_watermark = HOST_WIDE_INT_M1U; | |
3219 | } | |
3220 | ||
3221 | /* If we have no actual push instructions, or shouldn't use them, | |
3222 | make space for all args right now. */ | |
3223 | else if (adjusted_args_size.var != 0) | |
3224 | { | |
3225 | if (old_stack_level == 0) | |
3226 | { | |
3227 | emit_stack_save (SAVE_BLOCK, &old_stack_level); | |
3228 | old_stack_pointer_delta = stack_pointer_delta; | |
3229 | old_pending_adj = pending_stack_adjust; | |
3230 | pending_stack_adjust = 0; | |
3231 | /* stack_arg_under_construction says whether a stack arg is | |
3232 | being constructed at the old stack level. Pushing the stack | |
3233 | gets a clean outgoing argument block. */ | |
3234 | old_stack_arg_under_construction = stack_arg_under_construction; | |
3235 | stack_arg_under_construction = 0; | |
3236 | } | |
3237 | argblock = push_block (ARGS_SIZE_RTX (adjusted_args_size), 0, 0); | |
3238 | if (flag_stack_usage_info) | |
3239 | current_function_has_unbounded_dynamic_stack_size = 1; | |
3240 | } | |
3241 | else | |
3242 | { | |
3243 | /* Note that we must go through the motions of allocating an argument | |
3244 | block even if the size is zero because we may be storing args | |
3245 | in the area reserved for register arguments, which may be part of | |
3246 | the stack frame. */ | |
3247 | ||
3248 | poly_int64 needed = adjusted_args_size.constant; | |
3249 | ||
3250 | /* Store the maximum argument space used. It will be pushed by | |
3251 | the prologue (if ACCUMULATE_OUTGOING_ARGS, or stack overflow | |
3252 | checking). */ | |
3253 | ||
3254 | crtl->outgoing_args_size = upper_bound (crtl->outgoing_args_size, | |
3255 | needed); | |
3256 | ||
3257 | if (must_preallocate) | |
3258 | { | |
3259 | if (ACCUMULATE_OUTGOING_ARGS) | |
3260 | { | |
3261 | /* Since the stack pointer will never be pushed, it is | |
3262 | possible for the evaluation of a parm to clobber | |
3263 | something we have already written to the stack. | |
3264 | Since most function calls on RISC machines do not use | |
3265 | the stack, this is uncommon, but must work correctly. | |
3266 | ||
3267 | Therefore, we save any area of the stack that was already | |
3268 | written and that we are using. Here we set up to do this | |
3269 | by making a new stack usage map from the old one. The | |
3270 | actual save will be done by store_one_arg. | |
3271 | ||
3272 | Another approach might be to try to reorder the argument | |
3273 | evaluations to avoid this conflicting stack usage. */ | |
3274 | ||
3275 | /* Since we will be writing into the entire argument area, | |
3276 | the map must be allocated for its entire size, not just | |
3277 | the part that is the responsibility of the caller. */ | |
3278 | if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))) | |
3279 | needed += reg_parm_stack_space; | |
3280 | ||
3281 | poly_int64 limit = needed; | |
3282 | if (ARGS_GROW_DOWNWARD) | |
3283 | limit += 1; | |
3284 | ||
3285 | /* For polynomial sizes, this is the maximum possible | |
3286 | size needed for arguments with a constant size | |
3287 | and offset. */ | |
3288 | HOST_WIDE_INT const_limit = constant_lower_bound (limit); | |
3289 | highest_outgoing_arg_in_use | |
3290 | = MAX (initial_highest_arg_in_use, const_limit); | |
3291 | ||
3292 | free (stack_usage_map_buf); | |
3293 | stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use); | |
3294 | stack_usage_map = stack_usage_map_buf; | |
3295 | ||
3296 | if (initial_highest_arg_in_use) | |
3297 | memcpy (stack_usage_map, initial_stack_usage_map, | |
3298 | initial_highest_arg_in_use); | |
3299 | ||
3300 | if (initial_highest_arg_in_use != highest_outgoing_arg_in_use) | |
3301 | memset (&stack_usage_map[initial_highest_arg_in_use], 0, | |
3302 | (highest_outgoing_arg_in_use | |
3303 | - initial_highest_arg_in_use)); | |
3304 | needed = 0; | |
3305 | ||
3306 | /* The address of the outgoing argument list must not be | |
3307 | copied to a register here, because argblock would be left | |
3308 | pointing to the wrong place after the call to | |
3309 | allocate_dynamic_stack_space below. */ | |
3310 | ||
3311 | argblock = virtual_outgoing_args_rtx; | |
3312 | } | |
3313 | else | |
3314 | { | |
3315 | /* Try to reuse some or all of the pending_stack_adjust | |
3316 | to get this space. */ | |
3317 | if (inhibit_defer_pop == 0 | |
3318 | && (combine_pending_stack_adjustment_and_call | |
3319 | (&needed, | |
3320 | unadjusted_args_size, | |
3321 | &adjusted_args_size, | |
3322 | preferred_unit_stack_boundary))) | |
3323 | { | |
3324 | /* combine_pending_stack_adjustment_and_call computes | |
3325 | an adjustment before the arguments are allocated. | |
3326 | Account for them and see whether or not the stack | |
3327 | needs to go up or down. */ | |
3328 | needed = unadjusted_args_size - needed; | |
3329 | ||
3330 | /* Checked by | |
3331 | combine_pending_stack_adjustment_and_call. */ | |
3332 | gcc_checking_assert (ordered_p (needed, 0)); | |
3333 | if (maybe_lt (needed, 0)) | |
3334 | { | |
3335 | /* We're releasing stack space. */ | |
3336 | /* ??? We can avoid any adjustment at all if we're | |
3337 | already aligned. FIXME. */ | |
3338 | pending_stack_adjust = -needed; | |
3339 | do_pending_stack_adjust (); | |
3340 | needed = 0; | |
3341 | } | |
3342 | else | |
3343 | /* We need to allocate space. We'll do that in | |
3344 | push_block below. */ | |
3345 | pending_stack_adjust = 0; | |
3346 | } | |
3347 | ||
3348 | /* Special case this because overhead of `push_block' in | |
3349 | this case is non-trivial. */ | |
3350 | if (known_eq (needed, 0)) | |
3351 | argblock = virtual_outgoing_args_rtx; | |
3352 | else | |
3353 | { | |
3354 | rtx needed_rtx = gen_int_mode (needed, Pmode); | |
3355 | argblock = push_block (needed_rtx, 0, 0); | |
3356 | if (ARGS_GROW_DOWNWARD) | |
3357 | argblock = plus_constant (Pmode, argblock, needed); | |
3358 | } | |
3359 | ||
3360 | /* We only really need to call `copy_to_reg' in the case | |
3361 | where push insns are going to be used to pass ARGBLOCK | |
3362 | to a function call in ARGS. In that case, the stack | |
3363 | pointer changes value from the allocation point to the | |
3364 | call point, and hence the value of | |
3365 | VIRTUAL_OUTGOING_ARGS_RTX changes as well. But might | |
3366 | as well always do it. */ | |
3367 | argblock = copy_to_reg (argblock); | |
3368 | } | |
3369 | } | |
3370 | } | |
3371 | ||
3372 | if (ACCUMULATE_OUTGOING_ARGS) | |
3373 | { | |
3374 | /* The save/restore code in store_one_arg handles all | |
3375 | cases except one: a constructor call (including a C | |
3376 | function returning a BLKmode struct) to initialize | |
3377 | an argument. */ | |
3378 | if (stack_arg_under_construction) | |
3379 | { | |
3380 | rtx push_size | |
3381 | = (gen_int_mode | |
3382 | (adjusted_args_size.constant | |
3383 | + (OUTGOING_REG_PARM_STACK_SPACE (!fndecl ? fntype | |
3384 | : TREE_TYPE (fndecl)) | |
3385 | ? 0 : reg_parm_stack_space), Pmode)); | |
3386 | if (old_stack_level == 0) | |
3387 | { | |
3388 | emit_stack_save (SAVE_BLOCK, &old_stack_level); | |
3389 | old_stack_pointer_delta = stack_pointer_delta; | |
3390 | old_pending_adj = pending_stack_adjust; | |
3391 | pending_stack_adjust = 0; | |
3392 | /* stack_arg_under_construction says whether a stack | |
3393 | arg is being constructed at the old stack level. | |
3394 | Pushing the stack gets a clean outgoing argument | |
3395 | block. */ | |
3396 | old_stack_arg_under_construction | |
3397 | = stack_arg_under_construction; | |
3398 | stack_arg_under_construction = 0; | |
3399 | /* Make a new map for the new argument list. */ | |
3400 | free (stack_usage_map_buf); | |
3401 | stack_usage_map_buf = XCNEWVEC (char, highest_outgoing_arg_in_use); | |
3402 | stack_usage_map = stack_usage_map_buf; | |
3403 | highest_outgoing_arg_in_use = 0; | |
3404 | stack_usage_watermark = HOST_WIDE_INT_M1U; | |
3405 | } | |
3406 | /* We can pass TRUE as the 4th argument because we just | |
3407 | saved the stack pointer and will restore it right after | |
3408 | the call. */ | |
3409 | allocate_dynamic_stack_space (push_size, 0, BIGGEST_ALIGNMENT, | |
3410 | -1, true); | |
3411 | } | |
3412 | ||
3413 | /* If argument evaluation might modify the stack pointer, | |
3414 | copy the address of the argument list to a register. */ | |
3415 | for (i = 0; i < num_actuals; i++) | |
3416 | if (args[i].pass_on_stack) | |
3417 | { | |
3418 | argblock = copy_addr_to_reg (argblock); | |
3419 | break; | |
3420 | } | |
3421 | } | |
3422 | ||
3423 | compute_argument_addresses (args, argblock, num_actuals); | |
3424 | ||
3425 | /* Stack is properly aligned, pops can't safely be deferred during | |
3426 | the evaluation of the arguments. */ | |
3427 | NO_DEFER_POP; | |
3428 | ||
3429 | /* Precompute all register parameters. It isn't safe to compute | |
3430 | anything once we have started filling any specific hard regs. | |
3431 | TLS symbols sometimes need a call to resolve. Precompute | |
3432 | register parameters before any stack pointer manipulation | |
3433 | to avoid unaligned stack in the called function. */ | |
3434 | precompute_register_parameters (num_actuals, args, ®_parm_seen); | |
3435 | ||
3436 | OK_DEFER_POP; | |
3437 | ||
3438 | /* Perform stack alignment before the first push (the last arg). */ | |
3439 | if (argblock == 0 | |
3440 | && maybe_gt (adjusted_args_size.constant, reg_parm_stack_space) | |
3441 | && maybe_ne (adjusted_args_size.constant, unadjusted_args_size)) | |
3442 | { | |
3443 | /* When the stack adjustment is pending, we get better code | |
3444 | by combining the adjustments. */ | |
3445 | if (maybe_ne (pending_stack_adjust, 0) | |
3446 | && ! inhibit_defer_pop | |
3447 | && (combine_pending_stack_adjustment_and_call | |
3448 | (&pending_stack_adjust, | |
3449 | unadjusted_args_size, | |
3450 | &adjusted_args_size, | |
3451 | preferred_unit_stack_boundary))) | |
3452 | do_pending_stack_adjust (); | |
3453 | else if (argblock == 0) | |
3454 | anti_adjust_stack (gen_int_mode (adjusted_args_size.constant | |
3455 | - unadjusted_args_size, | |
3456 | Pmode)); | |
3457 | } | |
3458 | /* Now that the stack is properly aligned, pops can't safely | |
3459 | be deferred during the evaluation of the arguments. */ | |
3460 | NO_DEFER_POP; | |
3461 | ||
3462 | /* Record the maximum pushed stack space size. We need to delay | |
3463 | doing it this far to take into account the optimization done | |
3464 | by combine_pending_stack_adjustment_and_call. */ | |
3465 | if (flag_stack_usage_info | |
3466 | && !ACCUMULATE_OUTGOING_ARGS | |
3467 | && pass | |
3468 | && adjusted_args_size.var == 0) | |
3469 | { | |
3470 | poly_int64 pushed = (adjusted_args_size.constant | |
3471 | + pending_stack_adjust); | |
3472 | current_function_pushed_stack_size | |
3473 | = upper_bound (current_function_pushed_stack_size, pushed); | |
3474 | } | |
3475 | ||
3476 | funexp = rtx_for_function_call (fndecl, addr); | |
3477 | ||
3478 | if (CALL_EXPR_STATIC_CHAIN (exp)) | |
3479 | static_chain_value = expand_normal (CALL_EXPR_STATIC_CHAIN (exp)); | |
3480 | else | |
3481 | static_chain_value = 0; | |
3482 | ||
3483 | #ifdef REG_PARM_STACK_SPACE | |
3484 | /* Save the fixed argument area if it's part of the caller's frame and | |
3485 | is clobbered by argument setup for this call. */ | |
3486 | if (ACCUMULATE_OUTGOING_ARGS && pass) | |
3487 | save_area = save_fixed_argument_area (reg_parm_stack_space, argblock, | |
3488 | &low_to_save, &high_to_save); | |
3489 | #endif | |
3490 | ||
3491 | /* Now store (and compute if necessary) all non-register parms. | |
3492 | These come before register parms, since they can require block-moves, | |
3493 | which could clobber the registers used for register parms. | |
3494 | Parms which have partial registers are not stored here, | |
3495 | but we do preallocate space here if they want that. */ | |
3496 | ||
3497 | for (i = 0; i < num_actuals; i++) | |
3498 | { | |
3499 | if (args[i].reg == 0 || args[i].pass_on_stack) | |
3500 | { | |
3501 | rtx_insn *before_arg = get_last_insn (); | |
3502 | ||
3503 | /* We don't allow passing huge (> 2^30 B) arguments | |
3504 | by value. It would cause an overflow later on. */ | |
3505 | if (constant_lower_bound (adjusted_args_size.constant) | |
3506 | >= (1 << (HOST_BITS_PER_INT - 2))) | |
3507 | { | |
3508 | sorry ("passing too large argument on stack"); | |
3509 | /* Don't worry about stack clean-up. */ | |
3510 | if (pass == 0) | |
3511 | sibcall_failure = true; | |
3512 | else | |
3513 | normal_failure = true; | |
3514 | continue; | |
3515 | } | |
3516 | ||
3517 | if (store_one_arg (&args[i], argblock, flags, | |
3518 | adjusted_args_size.var != 0, | |
3519 | reg_parm_stack_space) | |
3520 | || (pass == 0 | |
3521 | && check_sibcall_argument_overlap (before_arg, | |
3522 | &args[i], true))) | |
3523 | sibcall_failure = true; | |
3524 | } | |
3525 | ||
3526 | if (args[i].stack) | |
3527 | call_fusage | |
3528 | = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[i].tree_value)), | |
3529 | gen_rtx_USE (VOIDmode, args[i].stack), | |
3530 | call_fusage); | |
3531 | } | |
3532 | ||
3533 | /* If we have a parm that is passed in registers but not in memory | |
3534 | and whose alignment does not permit a direct copy into registers, | |
3535 | make a group of pseudos that correspond to each register that we | |
3536 | will later fill. */ | |
3537 | if (STRICT_ALIGNMENT) | |
3538 | store_unaligned_arguments_into_pseudos (args, num_actuals); | |
3539 | ||
3540 | /* Now store any partially-in-registers parm. | |
3541 | This is the last place a block-move can happen. */ | |
3542 | if (reg_parm_seen) | |
3543 | for (i = 0; i < num_actuals; i++) | |
3544 | if (args[i].partial != 0 && ! args[i].pass_on_stack) | |
3545 | { | |
3546 | rtx_insn *before_arg = get_last_insn (); | |
3547 | ||
3548 | /* On targets with weird calling conventions (e.g. PA) it's | |
3549 | hard to ensure that all cases of argument overlap between | |
3550 | stack and registers work. Play it safe and bail out. */ | |
3551 | if (ARGS_GROW_DOWNWARD && !STACK_GROWS_DOWNWARD) | |
3552 | { | |
3553 | sibcall_failure = true; | |
3554 | break; | |
3555 | } | |
3556 | ||
3557 | if (store_one_arg (&args[i], argblock, flags, | |
3558 | adjusted_args_size.var != 0, | |
3559 | reg_parm_stack_space) | |
3560 | || (pass == 0 | |
3561 | && check_sibcall_argument_overlap (before_arg, | |
3562 | &args[i], true))) | |
3563 | sibcall_failure = true; | |
3564 | } | |
3565 | ||
3566 | /* Set up the next argument register. For sibling calls on machines | |
3567 | with register windows this should be the incoming register. */ | |
3568 | if (pass == 0) | |
3569 | next_arg_reg = targetm.calls.function_incoming_arg | |
3570 | (args_so_far, function_arg_info::end_marker ()); | |
3571 | else | |
3572 | next_arg_reg = targetm.calls.function_arg | |
3573 | (args_so_far, function_arg_info::end_marker ()); | |
3574 | ||
3575 | targetm.calls.start_call_args (args_so_far); | |
3576 | ||
3577 | bool any_regs = false; | |
3578 | for (i = 0; i < num_actuals; i++) | |
3579 | if (args[i].reg != NULL_RTX) | |
3580 | { | |
3581 | any_regs = true; | |
3582 | targetm.calls.call_args (args_so_far, args[i].reg, funtype); | |
3583 | } | |
3584 | if (!any_regs) | |
3585 | targetm.calls.call_args (args_so_far, pc_rtx, funtype); | |
3586 | ||
3587 | /* Figure out the register where the value, if any, will come back. */ | |
3588 | valreg = 0; | |
3589 | if (TYPE_MODE (rettype) != VOIDmode | |
3590 | && ! structure_value_addr) | |
3591 | { | |
3592 | if (pcc_struct_value) | |
3593 | valreg = hard_function_value (build_pointer_type (rettype), | |
3594 | fndecl, NULL, (pass == 0)); | |
3595 | else | |
3596 | valreg = hard_function_value (rettype, fndecl, fntype, | |
3597 | (pass == 0)); | |
3598 | ||
3599 | /* If VALREG is a PARALLEL whose first member has a zero | |
3600 | offset, use that. This is for targets such as m68k that | |
3601 | return the same value in multiple places. */ | |
3602 | if (GET_CODE (valreg) == PARALLEL) | |
3603 | { | |
3604 | rtx elem = XVECEXP (valreg, 0, 0); | |
3605 | rtx where = XEXP (elem, 0); | |
3606 | rtx offset = XEXP (elem, 1); | |
3607 | if (offset == const0_rtx | |
3608 | && GET_MODE (where) == GET_MODE (valreg)) | |
3609 | valreg = where; | |
3610 | } | |
3611 | } | |
3612 | ||
3613 | /* If register arguments require space on the stack and stack space | |
3614 | was not preallocated, allocate stack space here for arguments | |
3615 | passed in registers. */ | |
3616 | if (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))) | |
3617 | && !ACCUMULATE_OUTGOING_ARGS | |
3618 | && !must_preallocate && reg_parm_stack_space > 0) | |
3619 | anti_adjust_stack (GEN_INT (reg_parm_stack_space)); | |
3620 | ||
3621 | /* Pass the function the address in which to return a | |
3622 | structure value. */ | |
3623 | if (pass != 0 && structure_value_addr && ! structure_value_addr_parm) | |
3624 | { | |
3625 | structure_value_addr | |
3626 | = convert_memory_address (Pmode, structure_value_addr); | |
3627 | emit_move_insn (struct_value, | |
3628 | force_reg (Pmode, | |
3629 | force_operand (structure_value_addr, | |
3630 | NULL_RTX))); | |
3631 | ||
3632 | if (REG_P (struct_value)) | |
3633 | use_reg (&call_fusage, struct_value); | |
3634 | } | |
3635 | ||
3636 | after_args = get_last_insn (); | |
3637 | funexp = prepare_call_address (fndecl ? fndecl : fntype, funexp, | |
3638 | static_chain_value, &call_fusage, | |
3639 | reg_parm_seen, flags); | |
3640 | ||
3641 | load_register_parameters (args, num_actuals, &call_fusage, flags, | |
3642 | pass == 0, &sibcall_failure); | |
3643 | ||
3644 | /* Save a pointer to the last insn before the call, so that we can | |
3645 | later safely search backwards to find the CALL_INSN. */ | |
3646 | before_call = get_last_insn (); | |
3647 | ||
3648 | if (pass == 1 && (return_flags & ERF_RETURNS_ARG)) | |
3649 | { | |
3650 | int arg_nr = return_flags & ERF_RETURN_ARG_MASK; | |
3651 | arg_nr = num_actuals - arg_nr - 1; | |
3652 | if (arg_nr >= 0 | |
3653 | && arg_nr < num_actuals | |
3654 | && args[arg_nr].reg | |
3655 | && valreg | |
3656 | && REG_P (valreg) | |
3657 | && GET_MODE (args[arg_nr].reg) == GET_MODE (valreg)) | |
3658 | call_fusage | |
3659 | = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[arg_nr].tree_value)), | |
3660 | gen_rtx_SET (valreg, args[arg_nr].reg), | |
3661 | call_fusage); | |
3662 | } | |
3663 | /* All arguments and registers used for the call must be set up by | |
3664 | now! */ | |
3665 | ||
3666 | /* Stack must be properly aligned now. */ | |
3667 | gcc_assert (!pass | |
3668 | || multiple_p (stack_pointer_delta, | |
3669 | preferred_unit_stack_boundary)); | |
3670 | ||
3671 | /* Generate the actual call instruction. */ | |
3672 | emit_call_1 (funexp, exp, fndecl, funtype, unadjusted_args_size, | |
3673 | adjusted_args_size.constant, struct_value_size, | |
3674 | next_arg_reg, valreg, old_inhibit_defer_pop, call_fusage, | |
3675 | flags, args_so_far); | |
3676 | ||
3677 | if (flag_ipa_ra) | |
3678 | { | |
3679 | rtx_call_insn *last; | |
3680 | rtx datum = NULL_RTX; | |
3681 | if (fndecl != NULL_TREE) | |
3682 | { | |
3683 | datum = XEXP (DECL_RTL (fndecl), 0); | |
3684 | gcc_assert (datum != NULL_RTX | |
3685 | && GET_CODE (datum) == SYMBOL_REF); | |
3686 | } | |
3687 | last = last_call_insn (); | |
3688 | add_reg_note (last, REG_CALL_DECL, datum); | |
3689 | } | |
3690 | ||
3691 | /* If the call setup or the call itself overlaps with anything | |
3692 | of the argument setup we probably clobbered our call address. | |
3693 | In that case we can't do sibcalls. */ | |
3694 | if (pass == 0 | |
3695 | && check_sibcall_argument_overlap (after_args, 0, false)) | |
3696 | sibcall_failure = true; | |
3697 | ||
3698 | /* If a non-BLKmode value is returned at the most significant end | |
3699 | of a register, shift the register right by the appropriate amount | |
3700 | and update VALREG accordingly. BLKmode values are handled by the | |
3701 | group load/store machinery below. */ | |
3702 | if (!structure_value_addr | |
3703 | && !pcc_struct_value | |
3704 | && TYPE_MODE (rettype) != VOIDmode | |
3705 | && TYPE_MODE (rettype) != BLKmode | |
3706 | && REG_P (valreg) | |
3707 | && targetm.calls.return_in_msb (rettype)) | |
3708 | { | |
3709 | if (shift_return_value (TYPE_MODE (rettype), false, valreg)) | |
3710 | sibcall_failure = true; | |
3711 | valreg = gen_rtx_REG (TYPE_MODE (rettype), REGNO (valreg)); | |
3712 | } | |
3713 | ||
3714 | if (pass && (flags & ECF_MALLOC)) | |
3715 | { | |
3716 | rtx temp = gen_reg_rtx (GET_MODE (valreg)); | |
3717 | rtx_insn *last, *insns; | |
3718 | ||
3719 | /* The return value from a malloc-like function is a pointer. */ | |
3720 | if (TREE_CODE (rettype) == POINTER_TYPE) | |
3721 | mark_reg_pointer (temp, MALLOC_ABI_ALIGNMENT); | |
3722 | ||
3723 | emit_move_insn (temp, valreg); | |
3724 | ||
3725 | /* The return value from a malloc-like function cannot alias | |
3726 | anything else. */ | |
3727 | last = get_last_insn (); | |
3728 | add_reg_note (last, REG_NOALIAS, temp); | |
3729 | ||
3730 | /* Write out the sequence. */ | |
3731 | insns = get_insns (); | |
3732 | end_sequence (); | |
3733 | emit_insn (insns); | |
3734 | valreg = temp; | |
3735 | } | |
3736 | ||
3737 | /* For calls to `setjmp', etc., inform | |
3738 | function.cc:setjmp_warnings that it should complain if | |
3739 | nonvolatile values are live. For functions that cannot | |
3740 | return, inform flow that control does not fall through. */ | |
3741 | ||
3742 | if ((flags & ECF_NORETURN) || pass == 0) | |
3743 | { | |
3744 | /* The barrier must be emitted | |
3745 | immediately after the CALL_INSN. Some ports emit more | |
3746 | than just a CALL_INSN above, so we must search for it here. */ | |
3747 | ||
3748 | rtx_insn *last = get_last_insn (); | |
3749 | while (!CALL_P (last)) | |
3750 | { | |
3751 | last = PREV_INSN (last); | |
3752 | /* There was no CALL_INSN? */ | |
3753 | gcc_assert (last != before_call); | |
3754 | } | |
3755 | ||
3756 | emit_barrier_after (last); | |
3757 | ||
3758 | /* Stack adjustments after a noreturn call are dead code. | |
3759 | However when NO_DEFER_POP is in effect, we must preserve | |
3760 | stack_pointer_delta. */ | |
3761 | if (inhibit_defer_pop == 0) | |
3762 | { | |
3763 | stack_pointer_delta = old_stack_allocated; | |
3764 | pending_stack_adjust = 0; | |
3765 | } | |
3766 | } | |
3767 | ||
3768 | /* If value type not void, return an rtx for the value. */ | |
3769 | ||
3770 | if (TYPE_MODE (rettype) == VOIDmode | |
3771 | || ignore) | |
3772 | target = const0_rtx; | |
3773 | else if (structure_value_addr) | |
3774 | { | |
3775 | if (target == 0 || !MEM_P (target)) | |
3776 | { | |
3777 | target | |
3778 | = gen_rtx_MEM (TYPE_MODE (rettype), | |
3779 | memory_address (TYPE_MODE (rettype), | |
3780 | structure_value_addr)); | |
3781 | set_mem_attributes (target, rettype, 1); | |
3782 | } | |
3783 | } | |
3784 | else if (pcc_struct_value) | |
3785 | { | |
3786 | /* This is the special C++ case where we need to | |
3787 | know what the true target was. We take care to | |
3788 | never use this value more than once in one expression. */ | |
3789 | target = gen_rtx_MEM (TYPE_MODE (rettype), | |
3790 | copy_to_reg (valreg)); | |
3791 | set_mem_attributes (target, rettype, 1); | |
3792 | } | |
3793 | /* Handle calls that return values in multiple non-contiguous locations. | |
3794 | The Irix 6 ABI has examples of this. */ | |
3795 | else if (GET_CODE (valreg) == PARALLEL) | |
3796 | { | |
3797 | if (target == 0) | |
3798 | target = emit_group_move_into_temps (valreg); | |
3799 | else if (rtx_equal_p (target, valreg)) | |
3800 | ; | |
3801 | else if (GET_CODE (target) == PARALLEL) | |
3802 | /* Handle the result of a emit_group_move_into_temps | |
3803 | call in the previous pass. */ | |
3804 | emit_group_move (target, valreg); | |
3805 | else | |
3806 | emit_group_store (target, valreg, rettype, | |
3807 | int_size_in_bytes (rettype)); | |
3808 | } | |
3809 | else if (target | |
3810 | && GET_MODE (target) == TYPE_MODE (rettype) | |
3811 | && GET_MODE (target) == GET_MODE (valreg)) | |
3812 | { | |
3813 | bool may_overlap = false; | |
3814 | ||
3815 | /* We have to copy a return value in a CLASS_LIKELY_SPILLED hard | |
3816 | reg to a plain register. */ | |
3817 | if (!REG_P (target) || HARD_REGISTER_P (target)) | |
3818 | valreg = avoid_likely_spilled_reg (valreg); | |
3819 | ||
3820 | /* If TARGET is a MEM in the argument area, and we have | |
3821 | saved part of the argument area, then we can't store | |
3822 | directly into TARGET as it may get overwritten when we | |
3823 | restore the argument save area below. Don't work too | |
3824 | hard though and simply force TARGET to a register if it | |
3825 | is a MEM; the optimizer is quite likely to sort it out. */ | |
3826 | if (ACCUMULATE_OUTGOING_ARGS && pass && MEM_P (target)) | |
3827 | for (i = 0; i < num_actuals; i++) | |
3828 | if (args[i].save_area) | |
3829 | { | |
3830 | may_overlap = true; | |
3831 | break; | |
3832 | } | |
3833 | ||
3834 | if (may_overlap) | |
3835 | target = copy_to_reg (valreg); | |
3836 | else | |
3837 | { | |
3838 | /* TARGET and VALREG cannot be equal at this point | |
3839 | because the latter would not have | |
3840 | REG_FUNCTION_VALUE_P true, while the former would if | |
3841 | it were referring to the same register. | |
3842 | ||
3843 | If they refer to the same register, this move will be | |
3844 | a no-op, except when function inlining is being | |
3845 | done. */ | |
3846 | emit_move_insn (target, valreg); | |
3847 | ||
3848 | /* If we are setting a MEM, this code must be executed. | |
3849 | Since it is emitted after the call insn, sibcall | |
3850 | optimization cannot be performed in that case. */ | |
3851 | if (MEM_P (target)) | |
3852 | sibcall_failure = true; | |
3853 | } | |
3854 | } | |
3855 | else | |
3856 | target = copy_to_reg (avoid_likely_spilled_reg (valreg)); | |
3857 | ||
3858 | /* If we promoted this return value, make the proper SUBREG. | |
3859 | TARGET might be const0_rtx here, so be careful. */ | |
3860 | if (REG_P (target) | |
3861 | && TYPE_MODE (rettype) != BLKmode | |
3862 | && GET_MODE (target) != TYPE_MODE (rettype)) | |
3863 | { | |
3864 | tree type = rettype; | |
3865 | int unsignedp = TYPE_UNSIGNED (type); | |
3866 | machine_mode ret_mode = TYPE_MODE (type); | |
3867 | machine_mode pmode; | |
3868 | ||
3869 | /* Ensure we promote as expected, and get the new unsignedness. */ | |
3870 | pmode = promote_function_mode (type, ret_mode, &unsignedp, | |
3871 | funtype, 1); | |
3872 | gcc_assert (GET_MODE (target) == pmode); | |
3873 | ||
3874 | if (SCALAR_INT_MODE_P (pmode) | |
3875 | && SCALAR_FLOAT_MODE_P (ret_mode) | |
3876 | && known_gt (GET_MODE_SIZE (pmode), GET_MODE_SIZE (ret_mode))) | |
3877 | target = convert_wider_int_to_float (ret_mode, pmode, target); | |
3878 | else | |
3879 | { | |
3880 | target = gen_lowpart_SUBREG (ret_mode, target); | |
3881 | SUBREG_PROMOTED_VAR_P (target) = 1; | |
3882 | SUBREG_PROMOTED_SET (target, unsignedp); | |
3883 | } | |
3884 | } | |
3885 | ||
3886 | /* If size of args is variable or this was a constructor call for a stack | |
3887 | argument, restore saved stack-pointer value. */ | |
3888 | ||
3889 | if (old_stack_level) | |
3890 | { | |
3891 | rtx_insn *prev = get_last_insn (); | |
3892 | ||
3893 | emit_stack_restore (SAVE_BLOCK, old_stack_level); | |
3894 | stack_pointer_delta = old_stack_pointer_delta; | |
3895 | ||
3896 | fixup_args_size_notes (prev, get_last_insn (), stack_pointer_delta); | |
3897 | ||
3898 | pending_stack_adjust = old_pending_adj; | |
3899 | old_stack_allocated = stack_pointer_delta - pending_stack_adjust; | |
3900 | stack_arg_under_construction = old_stack_arg_under_construction; | |
3901 | highest_outgoing_arg_in_use = initial_highest_arg_in_use; | |
3902 | stack_usage_map = initial_stack_usage_map; | |
3903 | stack_usage_watermark = initial_stack_usage_watermark; | |
3904 | sibcall_failure = true; | |
3905 | } | |
3906 | else if (ACCUMULATE_OUTGOING_ARGS && pass) | |
3907 | { | |
3908 | #ifdef REG_PARM_STACK_SPACE | |
3909 | if (save_area) | |
3910 | restore_fixed_argument_area (save_area, argblock, | |
3911 | high_to_save, low_to_save); | |
3912 | #endif | |
3913 | ||
3914 | /* If we saved any argument areas, restore them. */ | |
3915 | for (i = 0; i < num_actuals; i++) | |
3916 | if (args[i].save_area) | |
3917 | { | |
3918 | machine_mode save_mode = GET_MODE (args[i].save_area); | |
3919 | rtx stack_area | |
3920 | = gen_rtx_MEM (save_mode, | |
3921 | memory_address (save_mode, | |
3922 | XEXP (args[i].stack_slot, 0))); | |
3923 | ||
3924 | if (save_mode != BLKmode) | |
3925 | emit_move_insn (stack_area, args[i].save_area); | |
3926 | else | |
3927 | emit_block_move (stack_area, args[i].save_area, | |
3928 | (gen_int_mode | |
3929 | (args[i].locate.size.constant, Pmode)), | |
3930 | BLOCK_OP_CALL_PARM); | |
3931 | } | |
3932 | ||
3933 | highest_outgoing_arg_in_use = initial_highest_arg_in_use; | |
3934 | stack_usage_map = initial_stack_usage_map; | |
3935 | stack_usage_watermark = initial_stack_usage_watermark; | |
3936 | } | |
3937 | ||
3938 | /* If this was alloca, record the new stack level. */ | |
3939 | if (flags & ECF_MAY_BE_ALLOCA) | |
3940 | record_new_stack_level (); | |
3941 | ||
3942 | /* Free up storage we no longer need. */ | |
3943 | for (i = 0; i < num_actuals; ++i) | |
3944 | free (args[i].aligned_regs); | |
3945 | ||
3946 | targetm.calls.end_call_args (args_so_far); | |
3947 | ||
3948 | insns = get_insns (); | |
3949 | end_sequence (); | |
3950 | ||
3951 | if (pass == 0) | |
3952 | { | |
3953 | tail_call_insns = insns; | |
3954 | ||
3955 | /* Restore the pending stack adjustment now that we have | |
3956 | finished generating the sibling call sequence. */ | |
3957 | ||
3958 | restore_pending_stack_adjust (&save); | |
3959 | ||
3960 | /* Prepare arg structure for next iteration. */ | |
3961 | for (i = 0; i < num_actuals; i++) | |
3962 | { | |
3963 | args[i].value = 0; | |
3964 | args[i].aligned_regs = 0; | |
3965 | args[i].stack = 0; | |
3966 | } | |
3967 | ||
3968 | sbitmap_free (stored_args_map); | |
3969 | internal_arg_pointer_exp_state.scan_start = NULL; | |
3970 | internal_arg_pointer_exp_state.cache.release (); | |
3971 | } | |
3972 | else | |
3973 | { | |
3974 | normal_call_insns = insns; | |
3975 | ||
3976 | /* Verify that we've deallocated all the stack we used. */ | |
3977 | gcc_assert ((flags & ECF_NORETURN) | |
3978 | || normal_failure | |
3979 | || known_eq (old_stack_allocated, | |
3980 | stack_pointer_delta | |
3981 | - pending_stack_adjust)); | |
3982 | if (normal_failure) | |
3983 | normal_call_insns = NULL; | |
3984 | } | |
3985 | ||
3986 | /* If something prevents making this a sibling call, | |
3987 | zero out the sequence. */ | |
3988 | if (sibcall_failure) | |
3989 | tail_call_insns = NULL; | |
3990 | else | |
3991 | break; | |
3992 | } | |
3993 | ||
3994 | /* If tail call production succeeded, we need to remove REG_EQUIV notes on | |
3995 | arguments too, as argument area is now clobbered by the call. */ | |
3996 | if (tail_call_insns) | |
3997 | { | |
3998 | emit_insn (tail_call_insns); | |
3999 | crtl->tail_call_emit = true; | |
4000 | } | |
4001 | else | |
4002 | { | |
4003 | emit_insn (normal_call_insns); | |
4004 | if (try_tail_call) | |
4005 | /* Ideally we'd emit a message for all of the ways that it could | |
4006 | have failed. */ | |
4007 | maybe_complain_about_tail_call (exp, "tail call production failed"); | |
4008 | } | |
4009 | ||
4010 | currently_expanding_call--; | |
4011 | ||
4012 | free (stack_usage_map_buf); | |
4013 | free (args); | |
4014 | return target; | |
4015 | } | |
4016 | ||
4017 | /* A sibling call sequence invalidates any REG_EQUIV notes made for | |
4018 | this function's incoming arguments. | |
4019 | ||
4020 | At the start of RTL generation we know the only REG_EQUIV notes | |
4021 | in the rtl chain are those for incoming arguments, so we can look | |
4022 | for REG_EQUIV notes between the start of the function and the | |
4023 | NOTE_INSN_FUNCTION_BEG. | |
4024 | ||
4025 | This is (slight) overkill. We could keep track of the highest | |
4026 | argument we clobber and be more selective in removing notes, but it | |
4027 | does not seem to be worth the effort. */ | |
4028 | ||
4029 | void | |
4030 | fixup_tail_calls (void) | |
4031 | { | |
4032 | rtx_insn *insn; | |
4033 | ||
4034 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
4035 | { | |
4036 | rtx note; | |
4037 | ||
4038 | /* There are never REG_EQUIV notes for the incoming arguments | |
4039 | after the NOTE_INSN_FUNCTION_BEG note, so stop if we see it. */ | |
4040 | if (NOTE_P (insn) | |
4041 | && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG) | |
4042 | break; | |
4043 | ||
4044 | note = find_reg_note (insn, REG_EQUIV, 0); | |
4045 | if (note) | |
4046 | remove_note (insn, note); | |
4047 | note = find_reg_note (insn, REG_EQUIV, 0); | |
4048 | gcc_assert (!note); | |
4049 | } | |
4050 | } | |
4051 | ||
4052 | /* Traverse a list of TYPES and expand all complex types into their | |
4053 | components. */ | |
4054 | static tree | |
4055 | split_complex_types (tree types) | |
4056 | { | |
4057 | tree p; | |
4058 | ||
4059 | /* Before allocating memory, check for the common case of no complex. */ | |
4060 | for (p = types; p; p = TREE_CHAIN (p)) | |
4061 | { | |
4062 | tree type = TREE_VALUE (p); | |
4063 | if (TREE_CODE (type) == COMPLEX_TYPE | |
4064 | && targetm.calls.split_complex_arg (type)) | |
4065 | goto found; | |
4066 | } | |
4067 | return types; | |
4068 | ||
4069 | found: | |
4070 | types = copy_list (types); | |
4071 | ||
4072 | for (p = types; p; p = TREE_CHAIN (p)) | |
4073 | { | |
4074 | tree complex_type = TREE_VALUE (p); | |
4075 | ||
4076 | if (TREE_CODE (complex_type) == COMPLEX_TYPE | |
4077 | && targetm.calls.split_complex_arg (complex_type)) | |
4078 | { | |
4079 | tree next, imag; | |
4080 | ||
4081 | /* Rewrite complex type with component type. */ | |
4082 | TREE_VALUE (p) = TREE_TYPE (complex_type); | |
4083 | next = TREE_CHAIN (p); | |
4084 | ||
4085 | /* Add another component type for the imaginary part. */ | |
4086 | imag = build_tree_list (NULL_TREE, TREE_VALUE (p)); | |
4087 | TREE_CHAIN (p) = imag; | |
4088 | TREE_CHAIN (imag) = next; | |
4089 | ||
4090 | /* Skip the newly created node. */ | |
4091 | p = TREE_CHAIN (p); | |
4092 | } | |
4093 | } | |
4094 | ||
4095 | return types; | |
4096 | } | |
4097 | \f | |
4098 | /* Output a library call to function ORGFUN (a SYMBOL_REF rtx) | |
4099 | for a value of mode OUTMODE, | |
4100 | with NARGS different arguments, passed as ARGS. | |
4101 | Store the return value if RETVAL is nonzero: store it in VALUE if | |
4102 | VALUE is nonnull, otherwise pick a convenient location. In either | |
4103 | case return the location of the stored value. | |
4104 | ||
4105 | FN_TYPE should be LCT_NORMAL for `normal' calls, LCT_CONST for | |
4106 | `const' calls, LCT_PURE for `pure' calls, or another LCT_ value for | |
4107 | other types of library calls. */ | |
4108 | ||
4109 | rtx | |
4110 | emit_library_call_value_1 (int retval, rtx orgfun, rtx value, | |
4111 | enum libcall_type fn_type, | |
4112 | machine_mode outmode, int nargs, rtx_mode_t *args) | |
4113 | { | |
4114 | /* Total size in bytes of all the stack-parms scanned so far. */ | |
4115 | struct args_size args_size; | |
4116 | /* Size of arguments before any adjustments (such as rounding). */ | |
4117 | struct args_size original_args_size; | |
4118 | int argnum; | |
4119 | rtx fun; | |
4120 | /* Todo, choose the correct decl type of orgfun. Sadly this information | |
4121 | isn't present here, so we default to native calling abi here. */ | |
4122 | tree fndecl ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */ | |
4123 | tree fntype ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */ | |
4124 | int count; | |
4125 | rtx argblock = 0; | |
4126 | CUMULATIVE_ARGS args_so_far_v; | |
4127 | cumulative_args_t args_so_far; | |
4128 | struct arg | |
4129 | { | |
4130 | rtx value; | |
4131 | machine_mode mode; | |
4132 | rtx reg; | |
4133 | int partial; | |
4134 | struct locate_and_pad_arg_data locate; | |
4135 | rtx save_area; | |
4136 | }; | |
4137 | struct arg *argvec; | |
4138 | int old_inhibit_defer_pop = inhibit_defer_pop; | |
4139 | rtx call_fusage = 0; | |
4140 | rtx mem_value = 0; | |
4141 | rtx valreg; | |
4142 | bool pcc_struct_value = false; | |
4143 | poly_int64 struct_value_size = 0; | |
4144 | int flags; | |
4145 | int reg_parm_stack_space = 0; | |
4146 | poly_int64 needed; | |
4147 | rtx_insn *before_call; | |
4148 | bool have_push_fusage; | |
4149 | tree tfom; /* type_for_mode (outmode, 0) */ | |
4150 | ||
4151 | #ifdef REG_PARM_STACK_SPACE | |
4152 | /* Define the boundary of the register parm stack space that needs to be | |
4153 | save, if any. */ | |
4154 | int low_to_save = 0, high_to_save = 0; | |
4155 | rtx save_area = 0; /* Place that it is saved. */ | |
4156 | #endif | |
4157 | ||
4158 | /* Size of the stack reserved for parameter registers. */ | |
4159 | unsigned int initial_highest_arg_in_use = highest_outgoing_arg_in_use; | |
4160 | char *initial_stack_usage_map = stack_usage_map; | |
4161 | unsigned HOST_WIDE_INT initial_stack_usage_watermark = stack_usage_watermark; | |
4162 | char *stack_usage_map_buf = NULL; | |
4163 | ||
4164 | rtx struct_value = targetm.calls.struct_value_rtx (0, 0); | |
4165 | ||
4166 | #ifdef REG_PARM_STACK_SPACE | |
4167 | reg_parm_stack_space = REG_PARM_STACK_SPACE ((tree) 0); | |
4168 | #endif | |
4169 | ||
4170 | /* By default, library functions cannot throw. */ | |
4171 | flags = ECF_NOTHROW; | |
4172 | ||
4173 | switch (fn_type) | |
4174 | { | |
4175 | case LCT_NORMAL: | |
4176 | break; | |
4177 | case LCT_CONST: | |
4178 | flags |= ECF_CONST; | |
4179 | break; | |
4180 | case LCT_PURE: | |
4181 | flags |= ECF_PURE; | |
4182 | break; | |
4183 | case LCT_NORETURN: | |
4184 | flags |= ECF_NORETURN; | |
4185 | break; | |
4186 | case LCT_THROW: | |
4187 | flags &= ~ECF_NOTHROW; | |
4188 | break; | |
4189 | case LCT_RETURNS_TWICE: | |
4190 | flags = ECF_RETURNS_TWICE; | |
4191 | break; | |
4192 | } | |
4193 | fun = orgfun; | |
4194 | ||
4195 | /* Ensure current function's preferred stack boundary is at least | |
4196 | what we need. */ | |
4197 | if (crtl->preferred_stack_boundary < PREFERRED_STACK_BOUNDARY) | |
4198 | crtl->preferred_stack_boundary = PREFERRED_STACK_BOUNDARY; | |
4199 | ||
4200 | /* If this kind of value comes back in memory, | |
4201 | decide where in memory it should come back. */ | |
4202 | if (outmode != VOIDmode) | |
4203 | { | |
4204 | tfom = lang_hooks.types.type_for_mode (outmode, 0); | |
4205 | if (aggregate_value_p (tfom, 0)) | |
4206 | { | |
4207 | #ifdef PCC_STATIC_STRUCT_RETURN | |
4208 | rtx pointer_reg | |
4209 | = hard_function_value (build_pointer_type (tfom), 0, 0, 0); | |
4210 | mem_value = gen_rtx_MEM (outmode, pointer_reg); | |
4211 | pcc_struct_value = true; | |
4212 | if (value == 0) | |
4213 | value = gen_reg_rtx (outmode); | |
4214 | #else /* not PCC_STATIC_STRUCT_RETURN */ | |
4215 | struct_value_size = GET_MODE_SIZE (outmode); | |
4216 | if (value != 0 && MEM_P (value)) | |
4217 | mem_value = value; | |
4218 | else | |
4219 | mem_value = assign_temp (tfom, 1, 1); | |
4220 | #endif | |
4221 | /* This call returns a big structure. */ | |
4222 | flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE); | |
4223 | } | |
4224 | } | |
4225 | else | |
4226 | tfom = void_type_node; | |
4227 | ||
4228 | /* ??? Unfinished: must pass the memory address as an argument. */ | |
4229 | ||
4230 | /* Copy all the libcall-arguments out of the varargs data | |
4231 | and into a vector ARGVEC. | |
4232 | ||
4233 | Compute how to pass each argument. We only support a very small subset | |
4234 | of the full argument passing conventions to limit complexity here since | |
4235 | library functions shouldn't have many args. */ | |
4236 | ||
4237 | argvec = XALLOCAVEC (struct arg, nargs + 1); | |
4238 | memset (argvec, 0, (nargs + 1) * sizeof (struct arg)); | |
4239 | ||
4240 | #ifdef INIT_CUMULATIVE_LIBCALL_ARGS | |
4241 | INIT_CUMULATIVE_LIBCALL_ARGS (args_so_far_v, outmode, fun); | |
4242 | #else | |
4243 | INIT_CUMULATIVE_ARGS (args_so_far_v, NULL_TREE, fun, 0, nargs); | |
4244 | #endif | |
4245 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
4246 | ||
4247 | args_size.constant = 0; | |
4248 | args_size.var = 0; | |
4249 | ||
4250 | count = 0; | |
4251 | ||
4252 | push_temp_slots (); | |
4253 | ||
4254 | /* If there's a structure value address to be passed, | |
4255 | either pass it in the special place, or pass it as an extra argument. */ | |
4256 | if (mem_value && struct_value == 0 && ! pcc_struct_value) | |
4257 | { | |
4258 | rtx addr = XEXP (mem_value, 0); | |
4259 | ||
4260 | nargs++; | |
4261 | ||
4262 | /* Make sure it is a reasonable operand for a move or push insn. */ | |
4263 | if (!REG_P (addr) && !MEM_P (addr) | |
4264 | && !(CONSTANT_P (addr) | |
4265 | && targetm.legitimate_constant_p (Pmode, addr))) | |
4266 | addr = force_operand (addr, NULL_RTX); | |
4267 | ||
4268 | argvec[count].value = addr; | |
4269 | argvec[count].mode = Pmode; | |
4270 | argvec[count].partial = 0; | |
4271 | ||
4272 | function_arg_info ptr_arg (Pmode, /*named=*/true); | |
4273 | argvec[count].reg = targetm.calls.function_arg (args_so_far, ptr_arg); | |
4274 | gcc_assert (targetm.calls.arg_partial_bytes (args_so_far, ptr_arg) == 0); | |
4275 | ||
4276 | locate_and_pad_parm (Pmode, NULL_TREE, | |
4277 | #ifdef STACK_PARMS_IN_REG_PARM_AREA | |
4278 | 1, | |
4279 | #else | |
4280 | argvec[count].reg != 0, | |
4281 | #endif | |
4282 | reg_parm_stack_space, 0, | |
4283 | NULL_TREE, &args_size, &argvec[count].locate); | |
4284 | ||
4285 | if (argvec[count].reg == 0 || argvec[count].partial != 0 | |
4286 | || reg_parm_stack_space > 0) | |
4287 | args_size.constant += argvec[count].locate.size.constant; | |
4288 | ||
4289 | targetm.calls.function_arg_advance (args_so_far, ptr_arg); | |
4290 | ||
4291 | count++; | |
4292 | } | |
4293 | ||
4294 | for (unsigned int i = 0; count < nargs; i++, count++) | |
4295 | { | |
4296 | rtx val = args[i].first; | |
4297 | function_arg_info arg (args[i].second, /*named=*/true); | |
4298 | int unsigned_p = 0; | |
4299 | ||
4300 | /* We cannot convert the arg value to the mode the library wants here; | |
4301 | must do it earlier where we know the signedness of the arg. */ | |
4302 | gcc_assert (arg.mode != BLKmode | |
4303 | && (GET_MODE (val) == arg.mode | |
4304 | || GET_MODE (val) == VOIDmode)); | |
4305 | ||
4306 | /* Make sure it is a reasonable operand for a move or push insn. */ | |
4307 | if (!REG_P (val) && !MEM_P (val) | |
4308 | && !(CONSTANT_P (val) | |
4309 | && targetm.legitimate_constant_p (arg.mode, val))) | |
4310 | val = force_operand (val, NULL_RTX); | |
4311 | ||
4312 | if (pass_by_reference (&args_so_far_v, arg)) | |
4313 | { | |
4314 | rtx slot; | |
4315 | int must_copy = !reference_callee_copied (&args_so_far_v, arg); | |
4316 | ||
4317 | /* If this was a CONST function, it is now PURE since it now | |
4318 | reads memory. */ | |
4319 | if (flags & ECF_CONST) | |
4320 | { | |
4321 | flags &= ~ECF_CONST; | |
4322 | flags |= ECF_PURE; | |
4323 | } | |
4324 | ||
4325 | if (MEM_P (val) && !must_copy) | |
4326 | { | |
4327 | tree val_expr = MEM_EXPR (val); | |
4328 | if (val_expr) | |
4329 | mark_addressable (val_expr); | |
4330 | slot = val; | |
4331 | } | |
4332 | else | |
4333 | { | |
4334 | slot = assign_temp (lang_hooks.types.type_for_mode (arg.mode, 0), | |
4335 | 1, 1); | |
4336 | emit_move_insn (slot, val); | |
4337 | } | |
4338 | ||
4339 | call_fusage = gen_rtx_EXPR_LIST (VOIDmode, | |
4340 | gen_rtx_USE (VOIDmode, slot), | |
4341 | call_fusage); | |
4342 | if (must_copy) | |
4343 | call_fusage = gen_rtx_EXPR_LIST (VOIDmode, | |
4344 | gen_rtx_CLOBBER (VOIDmode, | |
4345 | slot), | |
4346 | call_fusage); | |
4347 | ||
4348 | arg.mode = Pmode; | |
4349 | arg.pass_by_reference = true; | |
4350 | val = force_operand (XEXP (slot, 0), NULL_RTX); | |
4351 | } | |
4352 | ||
4353 | arg.mode = promote_function_mode (NULL_TREE, arg.mode, &unsigned_p, | |
4354 | NULL_TREE, 0); | |
4355 | argvec[count].mode = arg.mode; | |
4356 | argvec[count].value = convert_modes (arg.mode, GET_MODE (val), val, | |
4357 | unsigned_p); | |
4358 | argvec[count].reg = targetm.calls.function_arg (args_so_far, arg); | |
4359 | ||
4360 | argvec[count].partial | |
4361 | = targetm.calls.arg_partial_bytes (args_so_far, arg); | |
4362 | ||
4363 | if (argvec[count].reg == 0 | |
4364 | || argvec[count].partial != 0 | |
4365 | || reg_parm_stack_space > 0) | |
4366 | { | |
4367 | locate_and_pad_parm (arg.mode, NULL_TREE, | |
4368 | #ifdef STACK_PARMS_IN_REG_PARM_AREA | |
4369 | 1, | |
4370 | #else | |
4371 | argvec[count].reg != 0, | |
4372 | #endif | |
4373 | reg_parm_stack_space, argvec[count].partial, | |
4374 | NULL_TREE, &args_size, &argvec[count].locate); | |
4375 | args_size.constant += argvec[count].locate.size.constant; | |
4376 | gcc_assert (!argvec[count].locate.size.var); | |
4377 | } | |
4378 | #ifdef BLOCK_REG_PADDING | |
4379 | else | |
4380 | /* The argument is passed entirely in registers. See at which | |
4381 | end it should be padded. */ | |
4382 | argvec[count].locate.where_pad = | |
4383 | BLOCK_REG_PADDING (arg.mode, NULL_TREE, | |
4384 | known_le (GET_MODE_SIZE (arg.mode), | |
4385 | UNITS_PER_WORD)); | |
4386 | #endif | |
4387 | ||
4388 | targetm.calls.function_arg_advance (args_so_far, arg); | |
4389 | } | |
4390 | ||
4391 | for (int i = 0; i < nargs; i++) | |
4392 | if (reg_parm_stack_space > 0 | |
4393 | || argvec[i].reg == 0 | |
4394 | || argvec[i].partial != 0) | |
4395 | update_stack_alignment_for_call (&argvec[i].locate); | |
4396 | ||
4397 | /* If this machine requires an external definition for library | |
4398 | functions, write one out. */ | |
4399 | assemble_external_libcall (fun); | |
4400 | ||
4401 | original_args_size = args_size; | |
4402 | args_size.constant = (aligned_upper_bound (args_size.constant | |
4403 | + stack_pointer_delta, | |
4404 | STACK_BYTES) | |
4405 | - stack_pointer_delta); | |
4406 | ||
4407 | args_size.constant = upper_bound (args_size.constant, | |
4408 | reg_parm_stack_space); | |
4409 | ||
4410 | if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))) | |
4411 | args_size.constant -= reg_parm_stack_space; | |
4412 | ||
4413 | crtl->outgoing_args_size = upper_bound (crtl->outgoing_args_size, | |
4414 | args_size.constant); | |
4415 | ||
4416 | if (flag_stack_usage_info && !ACCUMULATE_OUTGOING_ARGS) | |
4417 | { | |
4418 | poly_int64 pushed = args_size.constant + pending_stack_adjust; | |
4419 | current_function_pushed_stack_size | |
4420 | = upper_bound (current_function_pushed_stack_size, pushed); | |
4421 | } | |
4422 | ||
4423 | if (ACCUMULATE_OUTGOING_ARGS) | |
4424 | { | |
4425 | /* Since the stack pointer will never be pushed, it is possible for | |
4426 | the evaluation of a parm to clobber something we have already | |
4427 | written to the stack. Since most function calls on RISC machines | |
4428 | do not use the stack, this is uncommon, but must work correctly. | |
4429 | ||
4430 | Therefore, we save any area of the stack that was already written | |
4431 | and that we are using. Here we set up to do this by making a new | |
4432 | stack usage map from the old one. | |
4433 | ||
4434 | Another approach might be to try to reorder the argument | |
4435 | evaluations to avoid this conflicting stack usage. */ | |
4436 | ||
4437 | needed = args_size.constant; | |
4438 | ||
4439 | /* Since we will be writing into the entire argument area, the | |
4440 | map must be allocated for its entire size, not just the part that | |
4441 | is the responsibility of the caller. */ | |
4442 | if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))) | |
4443 | needed += reg_parm_stack_space; | |
4444 | ||
4445 | poly_int64 limit = needed; | |
4446 | if (ARGS_GROW_DOWNWARD) | |
4447 | limit += 1; | |
4448 | ||
4449 | /* For polynomial sizes, this is the maximum possible size needed | |
4450 | for arguments with a constant size and offset. */ | |
4451 | HOST_WIDE_INT const_limit = constant_lower_bound (limit); | |
4452 | highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use, | |
4453 | const_limit); | |
4454 | ||
4455 | stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use); | |
4456 | stack_usage_map = stack_usage_map_buf; | |
4457 | ||
4458 | if (initial_highest_arg_in_use) | |
4459 | memcpy (stack_usage_map, initial_stack_usage_map, | |
4460 | initial_highest_arg_in_use); | |
4461 | ||
4462 | if (initial_highest_arg_in_use != highest_outgoing_arg_in_use) | |
4463 | memset (&stack_usage_map[initial_highest_arg_in_use], 0, | |
4464 | highest_outgoing_arg_in_use - initial_highest_arg_in_use); | |
4465 | needed = 0; | |
4466 | ||
4467 | /* We must be careful to use virtual regs before they're instantiated, | |
4468 | and real regs afterwards. Loop optimization, for example, can create | |
4469 | new libcalls after we've instantiated the virtual regs, and if we | |
4470 | use virtuals anyway, they won't match the rtl patterns. */ | |
4471 | ||
4472 | if (virtuals_instantiated) | |
4473 | argblock = plus_constant (Pmode, stack_pointer_rtx, | |
4474 | STACK_POINTER_OFFSET); | |
4475 | else | |
4476 | argblock = virtual_outgoing_args_rtx; | |
4477 | } | |
4478 | else | |
4479 | { | |
4480 | if (!targetm.calls.push_argument (0)) | |
4481 | argblock = push_block (gen_int_mode (args_size.constant, Pmode), 0, 0); | |
4482 | } | |
4483 | ||
4484 | /* We push args individually in reverse order, perform stack alignment | |
4485 | before the first push (the last arg). */ | |
4486 | if (argblock == 0) | |
4487 | anti_adjust_stack (gen_int_mode (args_size.constant | |
4488 | - original_args_size.constant, | |
4489 | Pmode)); | |
4490 | ||
4491 | argnum = nargs - 1; | |
4492 | ||
4493 | #ifdef REG_PARM_STACK_SPACE | |
4494 | if (ACCUMULATE_OUTGOING_ARGS) | |
4495 | { | |
4496 | /* The argument list is the property of the called routine and it | |
4497 | may clobber it. If the fixed area has been used for previous | |
4498 | parameters, we must save and restore it. */ | |
4499 | save_area = save_fixed_argument_area (reg_parm_stack_space, argblock, | |
4500 | &low_to_save, &high_to_save); | |
4501 | } | |
4502 | #endif | |
4503 | ||
4504 | rtx call_cookie | |
4505 | = targetm.calls.function_arg (args_so_far, | |
4506 | function_arg_info::end_marker ()); | |
4507 | ||
4508 | /* Push the args that need to be pushed. */ | |
4509 | ||
4510 | have_push_fusage = false; | |
4511 | ||
4512 | /* ARGNUM indexes the ARGVEC array in the order in which the arguments | |
4513 | are to be pushed. */ | |
4514 | for (count = 0; count < nargs; count++, argnum--) | |
4515 | { | |
4516 | machine_mode mode = argvec[argnum].mode; | |
4517 | rtx val = argvec[argnum].value; | |
4518 | rtx reg = argvec[argnum].reg; | |
4519 | int partial = argvec[argnum].partial; | |
4520 | unsigned int parm_align = argvec[argnum].locate.boundary; | |
4521 | poly_int64 lower_bound = 0, upper_bound = 0; | |
4522 | ||
4523 | if (! (reg != 0 && partial == 0)) | |
4524 | { | |
4525 | rtx use; | |
4526 | ||
4527 | if (ACCUMULATE_OUTGOING_ARGS) | |
4528 | { | |
4529 | /* If this is being stored into a pre-allocated, fixed-size, | |
4530 | stack area, save any previous data at that location. */ | |
4531 | ||
4532 | if (ARGS_GROW_DOWNWARD) | |
4533 | { | |
4534 | /* stack_slot is negative, but we want to index stack_usage_map | |
4535 | with positive values. */ | |
4536 | upper_bound = -argvec[argnum].locate.slot_offset.constant + 1; | |
4537 | lower_bound = upper_bound - argvec[argnum].locate.size.constant; | |
4538 | } | |
4539 | else | |
4540 | { | |
4541 | lower_bound = argvec[argnum].locate.slot_offset.constant; | |
4542 | upper_bound = lower_bound + argvec[argnum].locate.size.constant; | |
4543 | } | |
4544 | ||
4545 | if (stack_region_maybe_used_p (lower_bound, upper_bound, | |
4546 | reg_parm_stack_space)) | |
4547 | { | |
4548 | /* We need to make a save area. */ | |
4549 | poly_uint64 size | |
4550 | = argvec[argnum].locate.size.constant * BITS_PER_UNIT; | |
4551 | machine_mode save_mode | |
4552 | = int_mode_for_size (size, 1).else_blk (); | |
4553 | rtx adr | |
4554 | = plus_constant (Pmode, argblock, | |
4555 | argvec[argnum].locate.offset.constant); | |
4556 | rtx stack_area | |
4557 | = gen_rtx_MEM (save_mode, memory_address (save_mode, adr)); | |
4558 | ||
4559 | if (save_mode == BLKmode) | |
4560 | { | |
4561 | argvec[argnum].save_area | |
4562 | = assign_stack_temp (BLKmode, | |
4563 | argvec[argnum].locate.size.constant | |
4564 | ); | |
4565 | ||
4566 | emit_block_move (validize_mem | |
4567 | (copy_rtx (argvec[argnum].save_area)), | |
4568 | stack_area, | |
4569 | (gen_int_mode | |
4570 | (argvec[argnum].locate.size.constant, | |
4571 | Pmode)), | |
4572 | BLOCK_OP_CALL_PARM); | |
4573 | } | |
4574 | else | |
4575 | { | |
4576 | argvec[argnum].save_area = gen_reg_rtx (save_mode); | |
4577 | ||
4578 | emit_move_insn (argvec[argnum].save_area, stack_area); | |
4579 | } | |
4580 | } | |
4581 | } | |
4582 | ||
4583 | emit_push_insn (val, mode, NULL_TREE, NULL_RTX, parm_align, | |
4584 | partial, reg, 0, argblock, | |
4585 | (gen_int_mode | |
4586 | (argvec[argnum].locate.offset.constant, Pmode)), | |
4587 | reg_parm_stack_space, | |
4588 | ARGS_SIZE_RTX (argvec[argnum].locate.alignment_pad), false); | |
4589 | ||
4590 | /* Now mark the segment we just used. */ | |
4591 | if (ACCUMULATE_OUTGOING_ARGS) | |
4592 | mark_stack_region_used (lower_bound, upper_bound); | |
4593 | ||
4594 | NO_DEFER_POP; | |
4595 | ||
4596 | /* Indicate argument access so that alias.cc knows that these | |
4597 | values are live. */ | |
4598 | if (argblock) | |
4599 | use = plus_constant (Pmode, argblock, | |
4600 | argvec[argnum].locate.offset.constant); | |
4601 | else if (have_push_fusage) | |
4602 | continue; | |
4603 | else | |
4604 | { | |
4605 | /* When arguments are pushed, trying to tell alias.cc where | |
4606 | exactly this argument is won't work, because the | |
4607 | auto-increment causes confusion. So we merely indicate | |
4608 | that we access something with a known mode somewhere on | |
4609 | the stack. */ | |
4610 | use = gen_rtx_PLUS (Pmode, stack_pointer_rtx, | |
4611 | gen_rtx_SCRATCH (Pmode)); | |
4612 | have_push_fusage = true; | |
4613 | } | |
4614 | use = gen_rtx_MEM (argvec[argnum].mode, use); | |
4615 | use = gen_rtx_USE (VOIDmode, use); | |
4616 | call_fusage = gen_rtx_EXPR_LIST (VOIDmode, use, call_fusage); | |
4617 | } | |
4618 | } | |
4619 | ||
4620 | argnum = nargs - 1; | |
4621 | ||
4622 | fun = prepare_call_address (NULL, fun, NULL, &call_fusage, 0, 0); | |
4623 | ||
4624 | targetm.calls.start_call_args (args_so_far); | |
4625 | ||
4626 | /* When expanding a normal call, args are stored in push order, | |
4627 | which is the reverse of what we have here. */ | |
4628 | bool any_regs = false; | |
4629 | for (int i = nargs; i-- > 0; ) | |
4630 | if (argvec[i].reg != NULL_RTX) | |
4631 | { | |
4632 | targetm.calls.call_args (args_so_far, argvec[i].reg, NULL_TREE); | |
4633 | any_regs = true; | |
4634 | } | |
4635 | if (!any_regs) | |
4636 | targetm.calls.call_args (args_so_far, pc_rtx, NULL_TREE); | |
4637 | ||
4638 | /* Now load any reg parms into their regs. */ | |
4639 | ||
4640 | /* ARGNUM indexes the ARGVEC array in the order in which the arguments | |
4641 | are to be pushed. */ | |
4642 | for (count = 0; count < nargs; count++, argnum--) | |
4643 | { | |
4644 | machine_mode mode = argvec[argnum].mode; | |
4645 | rtx val = argvec[argnum].value; | |
4646 | rtx reg = argvec[argnum].reg; | |
4647 | int partial = argvec[argnum].partial; | |
4648 | ||
4649 | /* Handle calls that pass values in multiple non-contiguous | |
4650 | locations. The PA64 has examples of this for library calls. */ | |
4651 | if (reg != 0 && GET_CODE (reg) == PARALLEL) | |
4652 | emit_group_load (reg, val, NULL_TREE, GET_MODE_SIZE (mode)); | |
4653 | else if (reg != 0 && partial == 0) | |
4654 | { | |
4655 | emit_move_insn (reg, val); | |
4656 | #ifdef BLOCK_REG_PADDING | |
4657 | poly_int64 size = GET_MODE_SIZE (argvec[argnum].mode); | |
4658 | ||
4659 | /* Copied from load_register_parameters. */ | |
4660 | ||
4661 | /* Handle case where we have a value that needs shifting | |
4662 | up to the msb. eg. a QImode value and we're padding | |
4663 | upward on a BYTES_BIG_ENDIAN machine. */ | |
4664 | if (known_lt (size, UNITS_PER_WORD) | |
4665 | && (argvec[argnum].locate.where_pad | |
4666 | == (BYTES_BIG_ENDIAN ? PAD_UPWARD : PAD_DOWNWARD))) | |
4667 | { | |
4668 | rtx x; | |
4669 | poly_int64 shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT; | |
4670 | ||
4671 | /* Assigning REG here rather than a temp makes CALL_FUSAGE | |
4672 | report the whole reg as used. Strictly speaking, the | |
4673 | call only uses SIZE bytes at the msb end, but it doesn't | |
4674 | seem worth generating rtl to say that. */ | |
4675 | reg = gen_rtx_REG (word_mode, REGNO (reg)); | |
4676 | x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1); | |
4677 | if (x != reg) | |
4678 | emit_move_insn (reg, x); | |
4679 | } | |
4680 | #endif | |
4681 | } | |
4682 | ||
4683 | NO_DEFER_POP; | |
4684 | } | |
4685 | ||
4686 | /* Any regs containing parms remain in use through the call. */ | |
4687 | for (count = 0; count < nargs; count++) | |
4688 | { | |
4689 | rtx reg = argvec[count].reg; | |
4690 | if (reg != 0 && GET_CODE (reg) == PARALLEL) | |
4691 | use_group_regs (&call_fusage, reg); | |
4692 | else if (reg != 0) | |
4693 | { | |
4694 | int partial = argvec[count].partial; | |
4695 | if (partial) | |
4696 | { | |
4697 | int nregs; | |
4698 | gcc_assert (partial % UNITS_PER_WORD == 0); | |
4699 | nregs = partial / UNITS_PER_WORD; | |
4700 | use_regs (&call_fusage, REGNO (reg), nregs); | |
4701 | } | |
4702 | else | |
4703 | use_reg (&call_fusage, reg); | |
4704 | } | |
4705 | } | |
4706 | ||
4707 | /* Pass the function the address in which to return a structure value. */ | |
4708 | if (mem_value != 0 && struct_value != 0 && ! pcc_struct_value) | |
4709 | { | |
4710 | emit_move_insn (struct_value, | |
4711 | force_reg (Pmode, | |
4712 | force_operand (XEXP (mem_value, 0), | |
4713 | NULL_RTX))); | |
4714 | if (REG_P (struct_value)) | |
4715 | use_reg (&call_fusage, struct_value); | |
4716 | } | |
4717 | ||
4718 | /* Don't allow popping to be deferred, since then | |
4719 | cse'ing of library calls could delete a call and leave the pop. */ | |
4720 | NO_DEFER_POP; | |
4721 | valreg = (mem_value == 0 && outmode != VOIDmode | |
4722 | ? hard_libcall_value (outmode, orgfun) : NULL_RTX); | |
4723 | ||
4724 | /* Stack must be properly aligned now. */ | |
4725 | gcc_assert (multiple_p (stack_pointer_delta, | |
4726 | PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT)); | |
4727 | ||
4728 | before_call = get_last_insn (); | |
4729 | ||
4730 | if (flag_callgraph_info) | |
4731 | record_final_call (SYMBOL_REF_DECL (orgfun), UNKNOWN_LOCATION); | |
4732 | ||
4733 | /* We pass the old value of inhibit_defer_pop + 1 to emit_call_1, which | |
4734 | will set inhibit_defer_pop to that value. */ | |
4735 | /* The return type is needed to decide how many bytes the function pops. | |
4736 | Signedness plays no role in that, so for simplicity, we pretend it's | |
4737 | always signed. We also assume that the list of arguments passed has | |
4738 | no impact, so we pretend it is unknown. */ | |
4739 | ||
4740 | emit_call_1 (fun, NULL, | |
4741 | get_identifier (XSTR (orgfun, 0)), | |
4742 | build_function_type (tfom, NULL_TREE), | |
4743 | original_args_size.constant, args_size.constant, | |
4744 | struct_value_size, call_cookie, valreg, | |
4745 | old_inhibit_defer_pop + 1, call_fusage, flags, args_so_far); | |
4746 | ||
4747 | if (flag_ipa_ra) | |
4748 | { | |
4749 | rtx datum = orgfun; | |
4750 | gcc_assert (GET_CODE (datum) == SYMBOL_REF); | |
4751 | rtx_call_insn *last = last_call_insn (); | |
4752 | add_reg_note (last, REG_CALL_DECL, datum); | |
4753 | } | |
4754 | ||
4755 | /* Right-shift returned value if necessary. */ | |
4756 | if (!pcc_struct_value | |
4757 | && TYPE_MODE (tfom) != BLKmode | |
4758 | && targetm.calls.return_in_msb (tfom)) | |
4759 | { | |
4760 | shift_return_value (TYPE_MODE (tfom), false, valreg); | |
4761 | valreg = gen_rtx_REG (TYPE_MODE (tfom), REGNO (valreg)); | |
4762 | } | |
4763 | ||
4764 | targetm.calls.end_call_args (args_so_far); | |
4765 | ||
4766 | /* For calls to `setjmp', etc., inform function.cc:setjmp_warnings | |
4767 | that it should complain if nonvolatile values are live. For | |
4768 | functions that cannot return, inform flow that control does not | |
4769 | fall through. */ | |
4770 | if (flags & ECF_NORETURN) | |
4771 | { | |
4772 | /* The barrier note must be emitted | |
4773 | immediately after the CALL_INSN. Some ports emit more than | |
4774 | just a CALL_INSN above, so we must search for it here. */ | |
4775 | rtx_insn *last = get_last_insn (); | |
4776 | while (!CALL_P (last)) | |
4777 | { | |
4778 | last = PREV_INSN (last); | |
4779 | /* There was no CALL_INSN? */ | |
4780 | gcc_assert (last != before_call); | |
4781 | } | |
4782 | ||
4783 | emit_barrier_after (last); | |
4784 | } | |
4785 | ||
4786 | /* Consider that "regular" libcalls, i.e. all of them except for LCT_THROW | |
4787 | and LCT_RETURNS_TWICE, cannot perform non-local gotos. */ | |
4788 | if (flags & ECF_NOTHROW) | |
4789 | { | |
4790 | rtx_insn *last = get_last_insn (); | |
4791 | while (!CALL_P (last)) | |
4792 | { | |
4793 | last = PREV_INSN (last); | |
4794 | /* There was no CALL_INSN? */ | |
4795 | gcc_assert (last != before_call); | |
4796 | } | |
4797 | ||
4798 | make_reg_eh_region_note_nothrow_nononlocal (last); | |
4799 | } | |
4800 | ||
4801 | /* Now restore inhibit_defer_pop to its actual original value. */ | |
4802 | OK_DEFER_POP; | |
4803 | ||
4804 | pop_temp_slots (); | |
4805 | ||
4806 | /* Copy the value to the right place. */ | |
4807 | if (outmode != VOIDmode && retval) | |
4808 | { | |
4809 | if (mem_value) | |
4810 | { | |
4811 | if (value == 0) | |
4812 | value = mem_value; | |
4813 | if (value != mem_value) | |
4814 | emit_move_insn (value, mem_value); | |
4815 | } | |
4816 | else if (GET_CODE (valreg) == PARALLEL) | |
4817 | { | |
4818 | if (value == 0) | |
4819 | value = gen_reg_rtx (outmode); | |
4820 | emit_group_store (value, valreg, NULL_TREE, GET_MODE_SIZE (outmode)); | |
4821 | } | |
4822 | else | |
4823 | { | |
4824 | /* Convert to the proper mode if a promotion has been active. */ | |
4825 | if (GET_MODE (valreg) != outmode) | |
4826 | { | |
4827 | int unsignedp = TYPE_UNSIGNED (tfom); | |
4828 | ||
4829 | gcc_assert (promote_function_mode (tfom, outmode, &unsignedp, | |
4830 | fndecl ? TREE_TYPE (fndecl) : fntype, 1) | |
4831 | == GET_MODE (valreg)); | |
4832 | valreg = convert_modes (outmode, GET_MODE (valreg), valreg, 0); | |
4833 | } | |
4834 | ||
4835 | if (value != 0) | |
4836 | emit_move_insn (value, valreg); | |
4837 | else | |
4838 | value = valreg; | |
4839 | } | |
4840 | } | |
4841 | ||
4842 | if (ACCUMULATE_OUTGOING_ARGS) | |
4843 | { | |
4844 | #ifdef REG_PARM_STACK_SPACE | |
4845 | if (save_area) | |
4846 | restore_fixed_argument_area (save_area, argblock, | |
4847 | high_to_save, low_to_save); | |
4848 | #endif | |
4849 | ||
4850 | /* If we saved any argument areas, restore them. */ | |
4851 | for (count = 0; count < nargs; count++) | |
4852 | if (argvec[count].save_area) | |
4853 | { | |
4854 | machine_mode save_mode = GET_MODE (argvec[count].save_area); | |
4855 | rtx adr = plus_constant (Pmode, argblock, | |
4856 | argvec[count].locate.offset.constant); | |
4857 | rtx stack_area = gen_rtx_MEM (save_mode, | |
4858 | memory_address (save_mode, adr)); | |
4859 | ||
4860 | if (save_mode == BLKmode) | |
4861 | emit_block_move (stack_area, | |
4862 | validize_mem | |
4863 | (copy_rtx (argvec[count].save_area)), | |
4864 | (gen_int_mode | |
4865 | (argvec[count].locate.size.constant, Pmode)), | |
4866 | BLOCK_OP_CALL_PARM); | |
4867 | else | |
4868 | emit_move_insn (stack_area, argvec[count].save_area); | |
4869 | } | |
4870 | ||
4871 | highest_outgoing_arg_in_use = initial_highest_arg_in_use; | |
4872 | stack_usage_map = initial_stack_usage_map; | |
4873 | stack_usage_watermark = initial_stack_usage_watermark; | |
4874 | } | |
4875 | ||
4876 | free (stack_usage_map_buf); | |
4877 | ||
4878 | return value; | |
4879 | ||
4880 | } | |
4881 | \f | |
4882 | ||
4883 | /* Store a single argument for a function call | |
4884 | into the register or memory area where it must be passed. | |
4885 | *ARG describes the argument value and where to pass it. | |
4886 | ||
4887 | ARGBLOCK is the address of the stack-block for all the arguments, | |
4888 | or 0 on a machine where arguments are pushed individually. | |
4889 | ||
4890 | MAY_BE_ALLOCA nonzero says this could be a call to `alloca' | |
4891 | so must be careful about how the stack is used. | |
4892 | ||
4893 | VARIABLE_SIZE nonzero says that this was a variable-sized outgoing | |
4894 | argument stack. This is used if ACCUMULATE_OUTGOING_ARGS to indicate | |
4895 | that we need not worry about saving and restoring the stack. | |
4896 | ||
4897 | FNDECL is the declaration of the function we are calling. | |
4898 | ||
4899 | Return true if this arg should cause sibcall failure, | |
4900 | false otherwise. */ | |
4901 | ||
4902 | static bool | |
4903 | store_one_arg (struct arg_data *arg, rtx argblock, int flags, | |
4904 | int variable_size ATTRIBUTE_UNUSED, int reg_parm_stack_space) | |
4905 | { | |
4906 | tree pval = arg->tree_value; | |
4907 | rtx reg = 0; | |
4908 | int partial = 0; | |
4909 | poly_int64 used = 0; | |
4910 | poly_int64 lower_bound = 0, upper_bound = 0; | |
4911 | bool sibcall_failure = false; | |
4912 | ||
4913 | if (TREE_CODE (pval) == ERROR_MARK) | |
4914 | return true; | |
4915 | ||
4916 | /* Push a new temporary level for any temporaries we make for | |
4917 | this argument. */ | |
4918 | push_temp_slots (); | |
4919 | ||
4920 | if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL)) | |
4921 | { | |
4922 | /* If this is being stored into a pre-allocated, fixed-size, stack area, | |
4923 | save any previous data at that location. */ | |
4924 | if (argblock && ! variable_size && arg->stack) | |
4925 | { | |
4926 | if (ARGS_GROW_DOWNWARD) | |
4927 | { | |
4928 | /* stack_slot is negative, but we want to index stack_usage_map | |
4929 | with positive values. */ | |
4930 | if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS) | |
4931 | { | |
4932 | rtx offset = XEXP (XEXP (arg->stack_slot, 0), 1); | |
4933 | upper_bound = -rtx_to_poly_int64 (offset) + 1; | |
4934 | } | |
4935 | else | |
4936 | upper_bound = 0; | |
4937 | ||
4938 | lower_bound = upper_bound - arg->locate.size.constant; | |
4939 | } | |
4940 | else | |
4941 | { | |
4942 | if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS) | |
4943 | { | |
4944 | rtx offset = XEXP (XEXP (arg->stack_slot, 0), 1); | |
4945 | lower_bound = rtx_to_poly_int64 (offset); | |
4946 | } | |
4947 | else | |
4948 | lower_bound = 0; | |
4949 | ||
4950 | upper_bound = lower_bound + arg->locate.size.constant; | |
4951 | } | |
4952 | ||
4953 | if (stack_region_maybe_used_p (lower_bound, upper_bound, | |
4954 | reg_parm_stack_space)) | |
4955 | { | |
4956 | /* We need to make a save area. */ | |
4957 | poly_uint64 size = arg->locate.size.constant * BITS_PER_UNIT; | |
4958 | machine_mode save_mode | |
4959 | = int_mode_for_size (size, 1).else_blk (); | |
4960 | rtx adr = memory_address (save_mode, XEXP (arg->stack_slot, 0)); | |
4961 | rtx stack_area = gen_rtx_MEM (save_mode, adr); | |
4962 | ||
4963 | if (save_mode == BLKmode) | |
4964 | { | |
4965 | arg->save_area | |
4966 | = assign_temp (TREE_TYPE (arg->tree_value), 1, 1); | |
4967 | preserve_temp_slots (arg->save_area); | |
4968 | emit_block_move (validize_mem (copy_rtx (arg->save_area)), | |
4969 | stack_area, | |
4970 | (gen_int_mode | |
4971 | (arg->locate.size.constant, Pmode)), | |
4972 | BLOCK_OP_CALL_PARM); | |
4973 | } | |
4974 | else | |
4975 | { | |
4976 | arg->save_area = gen_reg_rtx (save_mode); | |
4977 | emit_move_insn (arg->save_area, stack_area); | |
4978 | } | |
4979 | } | |
4980 | } | |
4981 | } | |
4982 | ||
4983 | /* If this isn't going to be placed on both the stack and in registers, | |
4984 | set up the register and number of words. */ | |
4985 | if (! arg->pass_on_stack) | |
4986 | { | |
4987 | if (flags & ECF_SIBCALL) | |
4988 | reg = arg->tail_call_reg; | |
4989 | else | |
4990 | reg = arg->reg; | |
4991 | partial = arg->partial; | |
4992 | } | |
4993 | ||
4994 | /* Being passed entirely in a register. We shouldn't be called in | |
4995 | this case. */ | |
4996 | gcc_assert (reg == 0 || partial != 0); | |
4997 | ||
4998 | /* If this arg needs special alignment, don't load the registers | |
4999 | here. */ | |
5000 | if (arg->n_aligned_regs != 0) | |
5001 | reg = 0; | |
5002 | ||
5003 | /* If this is being passed partially in a register, we can't evaluate | |
5004 | it directly into its stack slot. Otherwise, we can. */ | |
5005 | if (arg->value == 0) | |
5006 | { | |
5007 | /* stack_arg_under_construction is nonzero if a function argument is | |
5008 | being evaluated directly into the outgoing argument list and | |
5009 | expand_call must take special action to preserve the argument list | |
5010 | if it is called recursively. | |
5011 | ||
5012 | For scalar function arguments stack_usage_map is sufficient to | |
5013 | determine which stack slots must be saved and restored. Scalar | |
5014 | arguments in general have pass_on_stack == false. | |
5015 | ||
5016 | If this argument is initialized by a function which takes the | |
5017 | address of the argument (a C++ constructor or a C function | |
5018 | returning a BLKmode structure), then stack_usage_map is | |
5019 | insufficient and expand_call must push the stack around the | |
5020 | function call. Such arguments have pass_on_stack == true. | |
5021 | ||
5022 | Note that it is always safe to set stack_arg_under_construction, | |
5023 | but this generates suboptimal code if set when not needed. */ | |
5024 | ||
5025 | if (arg->pass_on_stack) | |
5026 | stack_arg_under_construction++; | |
5027 | ||
5028 | arg->value = expand_expr (pval, | |
5029 | (partial | |
5030 | || TYPE_MODE (TREE_TYPE (pval)) != arg->mode) | |
5031 | ? NULL_RTX : arg->stack, | |
5032 | VOIDmode, EXPAND_STACK_PARM); | |
5033 | ||
5034 | /* If we are promoting object (or for any other reason) the mode | |
5035 | doesn't agree, convert the mode. */ | |
5036 | ||
5037 | if (arg->mode != TYPE_MODE (TREE_TYPE (pval))) | |
5038 | arg->value = convert_modes (arg->mode, TYPE_MODE (TREE_TYPE (pval)), | |
5039 | arg->value, arg->unsignedp); | |
5040 | ||
5041 | if (arg->pass_on_stack) | |
5042 | stack_arg_under_construction--; | |
5043 | } | |
5044 | ||
5045 | /* Check for overlap with already clobbered argument area. */ | |
5046 | if ((flags & ECF_SIBCALL) | |
5047 | && MEM_P (arg->value) | |
5048 | && mem_might_overlap_already_clobbered_arg_p (XEXP (arg->value, 0), | |
5049 | arg->locate.size.constant)) | |
5050 | sibcall_failure = true; | |
5051 | ||
5052 | /* Don't allow anything left on stack from computation | |
5053 | of argument to alloca. */ | |
5054 | if (flags & ECF_MAY_BE_ALLOCA) | |
5055 | do_pending_stack_adjust (); | |
5056 | ||
5057 | if (arg->value == arg->stack) | |
5058 | /* If the value is already in the stack slot, we are done. */ | |
5059 | ; | |
5060 | else if (arg->mode != BLKmode) | |
5061 | { | |
5062 | unsigned int parm_align; | |
5063 | ||
5064 | /* Argument is a scalar, not entirely passed in registers. | |
5065 | (If part is passed in registers, arg->partial says how much | |
5066 | and emit_push_insn will take care of putting it there.) | |
5067 | ||
5068 | Push it, and if its size is less than the | |
5069 | amount of space allocated to it, | |
5070 | also bump stack pointer by the additional space. | |
5071 | Note that in C the default argument promotions | |
5072 | will prevent such mismatches. */ | |
5073 | ||
5074 | poly_int64 size = (TYPE_EMPTY_P (TREE_TYPE (pval)) | |
5075 | ? 0 : GET_MODE_SIZE (arg->mode)); | |
5076 | ||
5077 | /* Compute how much space the push instruction will push. | |
5078 | On many machines, pushing a byte will advance the stack | |
5079 | pointer by a halfword. */ | |
5080 | #ifdef PUSH_ROUNDING | |
5081 | size = PUSH_ROUNDING (size); | |
5082 | #endif | |
5083 | used = size; | |
5084 | ||
5085 | /* Compute how much space the argument should get: | |
5086 | round up to a multiple of the alignment for arguments. */ | |
5087 | if (targetm.calls.function_arg_padding (arg->mode, TREE_TYPE (pval)) | |
5088 | != PAD_NONE) | |
5089 | /* At the moment we don't (need to) support ABIs for which the | |
5090 | padding isn't known at compile time. In principle it should | |
5091 | be easy to add though. */ | |
5092 | used = force_align_up (size, PARM_BOUNDARY / BITS_PER_UNIT); | |
5093 | ||
5094 | /* Compute the alignment of the pushed argument. */ | |
5095 | parm_align = arg->locate.boundary; | |
5096 | if (targetm.calls.function_arg_padding (arg->mode, TREE_TYPE (pval)) | |
5097 | == PAD_DOWNWARD) | |
5098 | { | |
5099 | poly_int64 pad = used - size; | |
5100 | unsigned int pad_align = known_alignment (pad) * BITS_PER_UNIT; | |
5101 | if (pad_align != 0) | |
5102 | parm_align = MIN (parm_align, pad_align); | |
5103 | } | |
5104 | ||
5105 | /* This isn't already where we want it on the stack, so put it there. | |
5106 | This can either be done with push or copy insns. */ | |
5107 | if (maybe_ne (used, 0) | |
5108 | && !emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), | |
5109 | NULL_RTX, parm_align, partial, reg, used - size, | |
5110 | argblock, ARGS_SIZE_RTX (arg->locate.offset), | |
5111 | reg_parm_stack_space, | |
5112 | ARGS_SIZE_RTX (arg->locate.alignment_pad), true)) | |
5113 | sibcall_failure = true; | |
5114 | ||
5115 | /* Unless this is a partially-in-register argument, the argument is now | |
5116 | in the stack. */ | |
5117 | if (partial == 0) | |
5118 | arg->value = arg->stack; | |
5119 | } | |
5120 | else | |
5121 | { | |
5122 | /* BLKmode, at least partly to be pushed. */ | |
5123 | ||
5124 | unsigned int parm_align; | |
5125 | poly_int64 excess; | |
5126 | rtx size_rtx; | |
5127 | ||
5128 | /* Pushing a nonscalar. | |
5129 | If part is passed in registers, PARTIAL says how much | |
5130 | and emit_push_insn will take care of putting it there. */ | |
5131 | ||
5132 | /* Round its size up to a multiple | |
5133 | of the allocation unit for arguments. */ | |
5134 | ||
5135 | if (arg->locate.size.var != 0) | |
5136 | { | |
5137 | excess = 0; | |
5138 | size_rtx = ARGS_SIZE_RTX (arg->locate.size); | |
5139 | } | |
5140 | else | |
5141 | { | |
5142 | /* PUSH_ROUNDING has no effect on us, because emit_push_insn | |
5143 | for BLKmode is careful to avoid it. */ | |
5144 | excess = (arg->locate.size.constant | |
5145 | - arg_int_size_in_bytes (TREE_TYPE (pval)) | |
5146 | + partial); | |
5147 | size_rtx = expand_expr (arg_size_in_bytes (TREE_TYPE (pval)), | |
5148 | NULL_RTX, TYPE_MODE (sizetype), | |
5149 | EXPAND_NORMAL); | |
5150 | } | |
5151 | ||
5152 | parm_align = arg->locate.boundary; | |
5153 | ||
5154 | /* When an argument is padded down, the block is aligned to | |
5155 | PARM_BOUNDARY, but the actual argument isn't. */ | |
5156 | if (targetm.calls.function_arg_padding (arg->mode, TREE_TYPE (pval)) | |
5157 | == PAD_DOWNWARD) | |
5158 | { | |
5159 | if (arg->locate.size.var) | |
5160 | parm_align = BITS_PER_UNIT; | |
5161 | else | |
5162 | { | |
5163 | unsigned int excess_align | |
5164 | = known_alignment (excess) * BITS_PER_UNIT; | |
5165 | if (excess_align != 0) | |
5166 | parm_align = MIN (parm_align, excess_align); | |
5167 | } | |
5168 | } | |
5169 | ||
5170 | if ((flags & ECF_SIBCALL) && MEM_P (arg->value)) | |
5171 | { | |
5172 | /* emit_push_insn might not work properly if arg->value and | |
5173 | argblock + arg->locate.offset areas overlap. */ | |
5174 | rtx x = arg->value; | |
5175 | poly_int64 i = 0; | |
5176 | ||
5177 | if (strip_offset (XEXP (x, 0), &i) | |
5178 | == crtl->args.internal_arg_pointer) | |
5179 | { | |
5180 | /* arg.locate doesn't contain the pretend_args_size offset, | |
5181 | it's part of argblock. Ensure we don't count it in I. */ | |
5182 | if (STACK_GROWS_DOWNWARD) | |
5183 | i -= crtl->args.pretend_args_size; | |
5184 | else | |
5185 | i += crtl->args.pretend_args_size; | |
5186 | ||
5187 | /* expand_call should ensure this. */ | |
5188 | gcc_assert (!arg->locate.offset.var | |
5189 | && arg->locate.size.var == 0); | |
5190 | poly_int64 size_val = rtx_to_poly_int64 (size_rtx); | |
5191 | ||
5192 | if (known_eq (arg->locate.offset.constant, i)) | |
5193 | { | |
5194 | /* Even though they appear to be at the same location, | |
5195 | if part of the outgoing argument is in registers, | |
5196 | they aren't really at the same location. Check for | |
5197 | this by making sure that the incoming size is the | |
5198 | same as the outgoing size. */ | |
5199 | if (maybe_ne (arg->locate.size.constant, size_val)) | |
5200 | sibcall_failure = true; | |
5201 | } | |
5202 | else if (maybe_in_range_p (arg->locate.offset.constant, | |
5203 | i, size_val)) | |
5204 | sibcall_failure = true; | |
5205 | /* Use arg->locate.size.constant instead of size_rtx | |
5206 | because we only care about the part of the argument | |
5207 | on the stack. */ | |
5208 | else if (maybe_in_range_p (i, arg->locate.offset.constant, | |
5209 | arg->locate.size.constant)) | |
5210 | sibcall_failure = true; | |
5211 | } | |
5212 | } | |
5213 | ||
5214 | if (!CONST_INT_P (size_rtx) || INTVAL (size_rtx) != 0) | |
5215 | emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), size_rtx, | |
5216 | parm_align, partial, reg, excess, argblock, | |
5217 | ARGS_SIZE_RTX (arg->locate.offset), | |
5218 | reg_parm_stack_space, | |
5219 | ARGS_SIZE_RTX (arg->locate.alignment_pad), false); | |
5220 | /* If we bypass emit_push_insn because it is a zero sized argument, | |
5221 | we still might need to adjust stack if such argument requires | |
5222 | extra alignment. See PR104558. */ | |
5223 | else if ((arg->locate.alignment_pad.var | |
5224 | || maybe_ne (arg->locate.alignment_pad.constant, 0)) | |
5225 | && !argblock) | |
5226 | anti_adjust_stack (ARGS_SIZE_RTX (arg->locate.alignment_pad)); | |
5227 | ||
5228 | /* Unless this is a partially-in-register argument, the argument is now | |
5229 | in the stack. | |
5230 | ||
5231 | ??? Unlike the case above, in which we want the actual | |
5232 | address of the data, so that we can load it directly into a | |
5233 | register, here we want the address of the stack slot, so that | |
5234 | it's properly aligned for word-by-word copying or something | |
5235 | like that. It's not clear that this is always correct. */ | |
5236 | if (partial == 0) | |
5237 | arg->value = arg->stack_slot; | |
5238 | } | |
5239 | ||
5240 | if (arg->reg && GET_CODE (arg->reg) == PARALLEL) | |
5241 | { | |
5242 | tree type = TREE_TYPE (arg->tree_value); | |
5243 | arg->parallel_value | |
5244 | = emit_group_load_into_temps (arg->reg, arg->value, type, | |
5245 | int_size_in_bytes (type)); | |
5246 | } | |
5247 | ||
5248 | /* Mark all slots this store used. */ | |
5249 | if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL) | |
5250 | && argblock && ! variable_size && arg->stack) | |
5251 | mark_stack_region_used (lower_bound, upper_bound); | |
5252 | ||
5253 | /* Once we have pushed something, pops can't safely | |
5254 | be deferred during the rest of the arguments. */ | |
5255 | NO_DEFER_POP; | |
5256 | ||
5257 | /* Free any temporary slots made in processing this argument. */ | |
5258 | pop_temp_slots (); | |
5259 | ||
5260 | return sibcall_failure; | |
5261 | } | |
5262 | ||
5263 | /* Nonzero if we do not know how to pass ARG solely in registers. */ | |
5264 | ||
5265 | bool | |
5266 | must_pass_in_stack_var_size (const function_arg_info &arg) | |
5267 | { | |
5268 | if (!arg.type) | |
5269 | return false; | |
5270 | ||
5271 | /* If the type has variable size... */ | |
5272 | if (!poly_int_tree_p (TYPE_SIZE (arg.type))) | |
5273 | return true; | |
5274 | ||
5275 | /* If the type is marked as addressable (it is required | |
5276 | to be constructed into the stack)... */ | |
5277 | if (TREE_ADDRESSABLE (arg.type)) | |
5278 | return true; | |
5279 | ||
5280 | return false; | |
5281 | } | |
5282 | ||
5283 | /* Another version of the TARGET_MUST_PASS_IN_STACK hook. This one | |
5284 | takes trailing padding of a structure into account. */ | |
5285 | /* ??? Should be able to merge these two by examining BLOCK_REG_PADDING. */ | |
5286 | ||
5287 | bool | |
5288 | must_pass_in_stack_var_size_or_pad (const function_arg_info &arg) | |
5289 | { | |
5290 | if (!arg.type) | |
5291 | return false; | |
5292 | ||
5293 | /* If the type has variable size... */ | |
5294 | if (TREE_CODE (TYPE_SIZE (arg.type)) != INTEGER_CST) | |
5295 | return true; | |
5296 | ||
5297 | /* If the type is marked as addressable (it is required | |
5298 | to be constructed into the stack)... */ | |
5299 | if (TREE_ADDRESSABLE (arg.type)) | |
5300 | return true; | |
5301 | ||
5302 | if (TYPE_EMPTY_P (arg.type)) | |
5303 | return false; | |
5304 | ||
5305 | /* If the padding and mode of the type is such that a copy into | |
5306 | a register would put it into the wrong part of the register. */ | |
5307 | if (arg.mode == BLKmode | |
5308 | && int_size_in_bytes (arg.type) % (PARM_BOUNDARY / BITS_PER_UNIT) | |
5309 | && (targetm.calls.function_arg_padding (arg.mode, arg.type) | |
5310 | == (BYTES_BIG_ENDIAN ? PAD_UPWARD : PAD_DOWNWARD))) | |
5311 | return true; | |
5312 | ||
5313 | return false; | |
5314 | } | |
5315 | ||
5316 | /* Return true if TYPE must be passed on the stack when passed to | |
5317 | the "..." arguments of a function. */ | |
5318 | ||
5319 | bool | |
5320 | must_pass_va_arg_in_stack (tree type) | |
5321 | { | |
5322 | function_arg_info arg (type, /*named=*/false); | |
5323 | return targetm.calls.must_pass_in_stack (arg); | |
5324 | } | |
5325 | ||
5326 | /* Return true if FIELD is the C++17 empty base field that should | |
5327 | be ignored for ABI calling convention decisions in order to | |
5328 | maintain ABI compatibility between C++14 and earlier, which doesn't | |
5329 | add this FIELD to classes with empty bases, and C++17 and later | |
5330 | which does. */ | |
5331 | ||
5332 | bool | |
5333 | cxx17_empty_base_field_p (const_tree field) | |
5334 | { | |
5335 | return (DECL_FIELD_ABI_IGNORED (field) | |
5336 | && DECL_ARTIFICIAL (field) | |
5337 | && RECORD_OR_UNION_TYPE_P (TREE_TYPE (field)) | |
5338 | && !lookup_attribute ("no_unique_address", DECL_ATTRIBUTES (field))); | |
5339 | } |