Mercurial > hg > CbC > CbC_gcc
comparison gcc/function.c @ 0:a06113de4d67
first commit
author | kent <kent@cr.ie.u-ryukyu.ac.jp> |
---|---|
date | Fri, 17 Jul 2009 14:47:48 +0900 |
parents | |
children | caeb520cebed 58ad6c70ea60 |
comparison
equal
deleted
inserted
replaced
-1:000000000000 | 0:a06113de4d67 |
---|---|
1 /* Expands front end tree to back end RTL for GCC. | |
2 Copyright (C) 1987, 1988, 1989, 1991, 1992, 1993, 1994, 1995, 1996, 1997, | |
3 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009 | |
4 Free Software Foundation, Inc. | |
5 | |
6 This file is part of GCC. | |
7 | |
8 GCC is free software; you can redistribute it and/or modify it under | |
9 the terms of the GNU General Public License as published by the Free | |
10 Software Foundation; either version 3, or (at your option) any later | |
11 version. | |
12 | |
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 for more details. | |
17 | |
18 You should have received a copy of the GNU General Public License | |
19 along with GCC; see the file COPYING3. If not see | |
20 <http://www.gnu.org/licenses/>. */ | |
21 | |
22 /* This file handles the generation of rtl code from tree structure | |
23 at the level of the function as a whole. | |
24 It creates the rtl expressions for parameters and auto variables | |
25 and has full responsibility for allocating stack slots. | |
26 | |
27 `expand_function_start' is called at the beginning of a function, | |
28 before the function body is parsed, and `expand_function_end' is | |
29 called after parsing the body. | |
30 | |
31 Call `assign_stack_local' to allocate a stack slot for a local variable. | |
32 This is usually done during the RTL generation for the function body, | |
33 but it can also be done in the reload pass when a pseudo-register does | |
34 not get a hard register. */ | |
35 | |
36 #include "config.h" | |
37 #include "system.h" | |
38 #include "coretypes.h" | |
39 #include "tm.h" | |
40 #include "rtl.h" | |
41 #include "tree.h" | |
42 #include "flags.h" | |
43 #include "except.h" | |
44 #include "function.h" | |
45 #include "expr.h" | |
46 #include "optabs.h" | |
47 #include "libfuncs.h" | |
48 #include "regs.h" | |
49 #include "hard-reg-set.h" | |
50 #include "insn-config.h" | |
51 #include "recog.h" | |
52 #include "output.h" | |
53 #include "basic-block.h" | |
54 #include "toplev.h" | |
55 #include "hashtab.h" | |
56 #include "ggc.h" | |
57 #include "tm_p.h" | |
58 #include "integrate.h" | |
59 #include "langhooks.h" | |
60 #include "target.h" | |
61 #include "cfglayout.h" | |
62 #include "gimple.h" | |
63 #include "tree-pass.h" | |
64 #include "predict.h" | |
65 #include "df.h" | |
66 #include "timevar.h" | |
67 #include "vecprim.h" | |
68 | |
69 /* So we can assign to cfun in this file. */ | |
70 #undef cfun | |
71 | |
72 #ifndef STACK_ALIGNMENT_NEEDED | |
73 #define STACK_ALIGNMENT_NEEDED 1 | |
74 #endif | |
75 | |
76 #define STACK_BYTES (STACK_BOUNDARY / BITS_PER_UNIT) | |
77 | |
78 /* Some systems use __main in a way incompatible with its use in gcc, in these | |
79 cases use the macros NAME__MAIN to give a quoted symbol and SYMBOL__MAIN to | |
80 give the same symbol without quotes for an alternative entry point. You | |
81 must define both, or neither. */ | |
82 #ifndef NAME__MAIN | |
83 #define NAME__MAIN "__main" | |
84 #endif | |
85 | |
86 /* Round a value to the lowest integer less than it that is a multiple of | |
87 the required alignment. Avoid using division in case the value is | |
88 negative. Assume the alignment is a power of two. */ | |
89 #define FLOOR_ROUND(VALUE,ALIGN) ((VALUE) & ~((ALIGN) - 1)) | |
90 | |
91 /* Similar, but round to the next highest integer that meets the | |
92 alignment. */ | |
93 #define CEIL_ROUND(VALUE,ALIGN) (((VALUE) + (ALIGN) - 1) & ~((ALIGN)- 1)) | |
94 | |
95 /* Nonzero if function being compiled doesn't contain any calls | |
96 (ignoring the prologue and epilogue). This is set prior to | |
97 local register allocation and is valid for the remaining | |
98 compiler passes. */ | |
99 int current_function_is_leaf; | |
100 | |
101 /* Nonzero if function being compiled doesn't modify the stack pointer | |
102 (ignoring the prologue and epilogue). This is only valid after | |
103 pass_stack_ptr_mod has run. */ | |
104 int current_function_sp_is_unchanging; | |
105 | |
106 /* Nonzero if the function being compiled is a leaf function which only | |
107 uses leaf registers. This is valid after reload (specifically after | |
108 sched2) and is useful only if the port defines LEAF_REGISTERS. */ | |
109 int current_function_uses_only_leaf_regs; | |
110 | |
111 /* Nonzero once virtual register instantiation has been done. | |
112 assign_stack_local uses frame_pointer_rtx when this is nonzero. | |
113 calls.c:emit_library_call_value_1 uses it to set up | |
114 post-instantiation libcalls. */ | |
115 int virtuals_instantiated; | |
116 | |
117 /* Assign unique numbers to labels generated for profiling, debugging, etc. */ | |
118 static GTY(()) int funcdef_no; | |
119 | |
120 /* These variables hold pointers to functions to create and destroy | |
121 target specific, per-function data structures. */ | |
122 struct machine_function * (*init_machine_status) (void); | |
123 | |
124 /* The currently compiled function. */ | |
125 struct function *cfun = 0; | |
126 | |
127 /* These arrays record the INSN_UIDs of the prologue and epilogue insns. */ | |
128 static VEC(int,heap) *prologue; | |
129 static VEC(int,heap) *epilogue; | |
130 | |
131 /* Array of INSN_UIDs to hold the INSN_UIDs for each sibcall epilogue | |
132 in this function. */ | |
133 static VEC(int,heap) *sibcall_epilogue; | |
134 | |
135 /* Forward declarations. */ | |
136 | |
137 static struct temp_slot *find_temp_slot_from_address (rtx); | |
138 static void pad_to_arg_alignment (struct args_size *, int, struct args_size *); | |
139 static void pad_below (struct args_size *, enum machine_mode, tree); | |
140 static void reorder_blocks_1 (rtx, tree, VEC(tree,heap) **); | |
141 static int all_blocks (tree, tree *); | |
142 static tree *get_block_vector (tree, int *); | |
143 extern tree debug_find_var_in_block_tree (tree, tree); | |
144 /* We always define `record_insns' even if it's not used so that we | |
145 can always export `prologue_epilogue_contains'. */ | |
146 static void record_insns (rtx, VEC(int,heap) **) ATTRIBUTE_UNUSED; | |
147 static int contains (const_rtx, VEC(int,heap) **); | |
148 #ifdef HAVE_return | |
149 static void emit_return_into_block (basic_block); | |
150 #endif | |
151 static void prepare_function_start (void); | |
152 static void do_clobber_return_reg (rtx, void *); | |
153 static void do_use_return_reg (rtx, void *); | |
154 static void set_insn_locators (rtx, int) ATTRIBUTE_UNUSED; | |
155 | |
156 /* Stack of nested functions. */ | |
157 /* Keep track of the cfun stack. */ | |
158 | |
159 typedef struct function *function_p; | |
160 | |
161 DEF_VEC_P(function_p); | |
162 DEF_VEC_ALLOC_P(function_p,heap); | |
163 static VEC(function_p,heap) *function_context_stack; | |
164 | |
165 /* Save the current context for compilation of a nested function. | |
166 This is called from language-specific code. */ | |
167 | |
168 void | |
169 push_function_context (void) | |
170 { | |
171 if (cfun == 0) | |
172 allocate_struct_function (NULL, false); | |
173 | |
174 VEC_safe_push (function_p, heap, function_context_stack, cfun); | |
175 set_cfun (NULL); | |
176 } | |
177 | |
178 /* Restore the last saved context, at the end of a nested function. | |
179 This function is called from language-specific code. */ | |
180 | |
181 void | |
182 pop_function_context (void) | |
183 { | |
184 struct function *p = VEC_pop (function_p, function_context_stack); | |
185 set_cfun (p); | |
186 current_function_decl = p->decl; | |
187 | |
188 /* Reset variables that have known state during rtx generation. */ | |
189 virtuals_instantiated = 0; | |
190 generating_concat_p = 1; | |
191 } | |
192 | |
193 /* Clear out all parts of the state in F that can safely be discarded | |
194 after the function has been parsed, but not compiled, to let | |
195 garbage collection reclaim the memory. */ | |
196 | |
197 void | |
198 free_after_parsing (struct function *f) | |
199 { | |
200 f->language = 0; | |
201 } | |
202 | |
203 /* Clear out all parts of the state in F that can safely be discarded | |
204 after the function has been compiled, to let garbage collection | |
205 reclaim the memory. */ | |
206 | |
207 void | |
208 free_after_compilation (struct function *f) | |
209 { | |
210 VEC_free (int, heap, prologue); | |
211 VEC_free (int, heap, epilogue); | |
212 VEC_free (int, heap, sibcall_epilogue); | |
213 if (crtl->emit.regno_pointer_align) | |
214 free (crtl->emit.regno_pointer_align); | |
215 | |
216 memset (crtl, 0, sizeof (struct rtl_data)); | |
217 f->eh = NULL; | |
218 f->machine = NULL; | |
219 f->cfg = NULL; | |
220 | |
221 regno_reg_rtx = NULL; | |
222 insn_locators_free (); | |
223 } | |
224 | |
225 /* Return size needed for stack frame based on slots so far allocated. | |
226 This size counts from zero. It is not rounded to PREFERRED_STACK_BOUNDARY; | |
227 the caller may have to do that. */ | |
228 | |
229 HOST_WIDE_INT | |
230 get_frame_size (void) | |
231 { | |
232 if (FRAME_GROWS_DOWNWARD) | |
233 return -frame_offset; | |
234 else | |
235 return frame_offset; | |
236 } | |
237 | |
238 /* Issue an error message and return TRUE if frame OFFSET overflows in | |
239 the signed target pointer arithmetics for function FUNC. Otherwise | |
240 return FALSE. */ | |
241 | |
242 bool | |
243 frame_offset_overflow (HOST_WIDE_INT offset, tree func) | |
244 { | |
245 unsigned HOST_WIDE_INT size = FRAME_GROWS_DOWNWARD ? -offset : offset; | |
246 | |
247 if (size > ((unsigned HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (Pmode) - 1)) | |
248 /* Leave room for the fixed part of the frame. */ | |
249 - 64 * UNITS_PER_WORD) | |
250 { | |
251 error ("%Jtotal size of local objects too large", func); | |
252 return TRUE; | |
253 } | |
254 | |
255 return FALSE; | |
256 } | |
257 | |
258 /* Return stack slot alignment in bits for TYPE and MODE. */ | |
259 | |
260 static unsigned int | |
261 get_stack_local_alignment (tree type, enum machine_mode mode) | |
262 { | |
263 unsigned int alignment; | |
264 | |
265 if (mode == BLKmode) | |
266 alignment = BIGGEST_ALIGNMENT; | |
267 else | |
268 alignment = GET_MODE_ALIGNMENT (mode); | |
269 | |
270 /* Allow the frond-end to (possibly) increase the alignment of this | |
271 stack slot. */ | |
272 if (! type) | |
273 type = lang_hooks.types.type_for_mode (mode, 0); | |
274 | |
275 return STACK_SLOT_ALIGNMENT (type, mode, alignment); | |
276 } | |
277 | |
278 /* Allocate a stack slot of SIZE bytes and return a MEM rtx for it | |
279 with machine mode MODE. | |
280 | |
281 ALIGN controls the amount of alignment for the address of the slot: | |
282 0 means according to MODE, | |
283 -1 means use BIGGEST_ALIGNMENT and round size to multiple of that, | |
284 -2 means use BITS_PER_UNIT, | |
285 positive specifies alignment boundary in bits. | |
286 | |
287 If REDUCE_ALIGNMENT_OK is true, it is OK to reduce alignment. | |
288 | |
289 We do not round to stack_boundary here. */ | |
290 | |
291 rtx | |
292 assign_stack_local_1 (enum machine_mode mode, HOST_WIDE_INT size, | |
293 int align, | |
294 bool reduce_alignment_ok ATTRIBUTE_UNUSED) | |
295 { | |
296 rtx x, addr; | |
297 int bigend_correction = 0; | |
298 unsigned int alignment, alignment_in_bits; | |
299 int frame_off, frame_alignment, frame_phase; | |
300 | |
301 if (align == 0) | |
302 { | |
303 alignment = get_stack_local_alignment (NULL, mode); | |
304 alignment /= BITS_PER_UNIT; | |
305 } | |
306 else if (align == -1) | |
307 { | |
308 alignment = BIGGEST_ALIGNMENT / BITS_PER_UNIT; | |
309 size = CEIL_ROUND (size, alignment); | |
310 } | |
311 else if (align == -2) | |
312 alignment = 1; /* BITS_PER_UNIT / BITS_PER_UNIT */ | |
313 else | |
314 alignment = align / BITS_PER_UNIT; | |
315 | |
316 alignment_in_bits = alignment * BITS_PER_UNIT; | |
317 | |
318 if (FRAME_GROWS_DOWNWARD) | |
319 frame_offset -= size; | |
320 | |
321 /* Ignore alignment if it exceeds MAX_SUPPORTED_STACK_ALIGNMENT. */ | |
322 if (alignment_in_bits > MAX_SUPPORTED_STACK_ALIGNMENT) | |
323 { | |
324 alignment_in_bits = MAX_SUPPORTED_STACK_ALIGNMENT; | |
325 alignment = alignment_in_bits / BITS_PER_UNIT; | |
326 } | |
327 | |
328 if (SUPPORTS_STACK_ALIGNMENT) | |
329 { | |
330 if (crtl->stack_alignment_estimated < alignment_in_bits) | |
331 { | |
332 if (!crtl->stack_realign_processed) | |
333 crtl->stack_alignment_estimated = alignment_in_bits; | |
334 else | |
335 { | |
336 /* If stack is realigned and stack alignment value | |
337 hasn't been finalized, it is OK not to increase | |
338 stack_alignment_estimated. The bigger alignment | |
339 requirement is recorded in stack_alignment_needed | |
340 below. */ | |
341 gcc_assert (!crtl->stack_realign_finalized); | |
342 if (!crtl->stack_realign_needed) | |
343 { | |
344 /* It is OK to reduce the alignment as long as the | |
345 requested size is 0 or the estimated stack | |
346 alignment >= mode alignment. */ | |
347 gcc_assert (reduce_alignment_ok | |
348 || size == 0 | |
349 || (crtl->stack_alignment_estimated | |
350 >= GET_MODE_ALIGNMENT (mode))); | |
351 alignment_in_bits = crtl->stack_alignment_estimated; | |
352 alignment = alignment_in_bits / BITS_PER_UNIT; | |
353 } | |
354 } | |
355 } | |
356 } | |
357 | |
358 if (crtl->stack_alignment_needed < alignment_in_bits) | |
359 crtl->stack_alignment_needed = alignment_in_bits; | |
360 if (crtl->max_used_stack_slot_alignment < crtl->stack_alignment_needed) | |
361 crtl->max_used_stack_slot_alignment = crtl->stack_alignment_needed; | |
362 | |
363 /* Calculate how many bytes the start of local variables is off from | |
364 stack alignment. */ | |
365 frame_alignment = PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT; | |
366 frame_off = STARTING_FRAME_OFFSET % frame_alignment; | |
367 frame_phase = frame_off ? frame_alignment - frame_off : 0; | |
368 | |
369 /* Round the frame offset to the specified alignment. The default is | |
370 to always honor requests to align the stack but a port may choose to | |
371 do its own stack alignment by defining STACK_ALIGNMENT_NEEDED. */ | |
372 if (STACK_ALIGNMENT_NEEDED | |
373 || mode != BLKmode | |
374 || size != 0) | |
375 { | |
376 /* We must be careful here, since FRAME_OFFSET might be negative and | |
377 division with a negative dividend isn't as well defined as we might | |
378 like. So we instead assume that ALIGNMENT is a power of two and | |
379 use logical operations which are unambiguous. */ | |
380 if (FRAME_GROWS_DOWNWARD) | |
381 frame_offset | |
382 = (FLOOR_ROUND (frame_offset - frame_phase, | |
383 (unsigned HOST_WIDE_INT) alignment) | |
384 + frame_phase); | |
385 else | |
386 frame_offset | |
387 = (CEIL_ROUND (frame_offset - frame_phase, | |
388 (unsigned HOST_WIDE_INT) alignment) | |
389 + frame_phase); | |
390 } | |
391 | |
392 /* On a big-endian machine, if we are allocating more space than we will use, | |
393 use the least significant bytes of those that are allocated. */ | |
394 if (BYTES_BIG_ENDIAN && mode != BLKmode && GET_MODE_SIZE (mode) < size) | |
395 bigend_correction = size - GET_MODE_SIZE (mode); | |
396 | |
397 /* If we have already instantiated virtual registers, return the actual | |
398 address relative to the frame pointer. */ | |
399 if (virtuals_instantiated) | |
400 addr = plus_constant (frame_pointer_rtx, | |
401 trunc_int_for_mode | |
402 (frame_offset + bigend_correction | |
403 + STARTING_FRAME_OFFSET, Pmode)); | |
404 else | |
405 addr = plus_constant (virtual_stack_vars_rtx, | |
406 trunc_int_for_mode | |
407 (frame_offset + bigend_correction, | |
408 Pmode)); | |
409 | |
410 if (!FRAME_GROWS_DOWNWARD) | |
411 frame_offset += size; | |
412 | |
413 x = gen_rtx_MEM (mode, addr); | |
414 set_mem_align (x, alignment_in_bits); | |
415 MEM_NOTRAP_P (x) = 1; | |
416 | |
417 stack_slot_list | |
418 = gen_rtx_EXPR_LIST (VOIDmode, x, stack_slot_list); | |
419 | |
420 if (frame_offset_overflow (frame_offset, current_function_decl)) | |
421 frame_offset = 0; | |
422 | |
423 return x; | |
424 } | |
425 | |
426 /* Wrap up assign_stack_local_1 with last parameter as false. */ | |
427 | |
428 rtx | |
429 assign_stack_local (enum machine_mode mode, HOST_WIDE_INT size, int align) | |
430 { | |
431 return assign_stack_local_1 (mode, size, align, false); | |
432 } | |
433 | |
434 | |
435 /* In order to evaluate some expressions, such as function calls returning | |
436 structures in memory, we need to temporarily allocate stack locations. | |
437 We record each allocated temporary in the following structure. | |
438 | |
439 Associated with each temporary slot is a nesting level. When we pop up | |
440 one level, all temporaries associated with the previous level are freed. | |
441 Normally, all temporaries are freed after the execution of the statement | |
442 in which they were created. However, if we are inside a ({...}) grouping, | |
443 the result may be in a temporary and hence must be preserved. If the | |
444 result could be in a temporary, we preserve it if we can determine which | |
445 one it is in. If we cannot determine which temporary may contain the | |
446 result, all temporaries are preserved. A temporary is preserved by | |
447 pretending it was allocated at the previous nesting level. | |
448 | |
449 Automatic variables are also assigned temporary slots, at the nesting | |
450 level where they are defined. They are marked a "kept" so that | |
451 free_temp_slots will not free them. */ | |
452 | |
453 struct temp_slot GTY(()) | |
454 { | |
455 /* Points to next temporary slot. */ | |
456 struct temp_slot *next; | |
457 /* Points to previous temporary slot. */ | |
458 struct temp_slot *prev; | |
459 /* The rtx to used to reference the slot. */ | |
460 rtx slot; | |
461 /* The size, in units, of the slot. */ | |
462 HOST_WIDE_INT size; | |
463 /* The type of the object in the slot, or zero if it doesn't correspond | |
464 to a type. We use this to determine whether a slot can be reused. | |
465 It can be reused if objects of the type of the new slot will always | |
466 conflict with objects of the type of the old slot. */ | |
467 tree type; | |
468 /* The alignment (in bits) of the slot. */ | |
469 unsigned int align; | |
470 /* Nonzero if this temporary is currently in use. */ | |
471 char in_use; | |
472 /* Nonzero if this temporary has its address taken. */ | |
473 char addr_taken; | |
474 /* Nesting level at which this slot is being used. */ | |
475 int level; | |
476 /* Nonzero if this should survive a call to free_temp_slots. */ | |
477 int keep; | |
478 /* The offset of the slot from the frame_pointer, including extra space | |
479 for alignment. This info is for combine_temp_slots. */ | |
480 HOST_WIDE_INT base_offset; | |
481 /* The size of the slot, including extra space for alignment. This | |
482 info is for combine_temp_slots. */ | |
483 HOST_WIDE_INT full_size; | |
484 }; | |
485 | |
486 /* A table of addresses that represent a stack slot. The table is a mapping | |
487 from address RTXen to a temp slot. */ | |
488 static GTY((param_is(struct temp_slot_address_entry))) htab_t temp_slot_address_table; | |
489 | |
490 /* Entry for the above hash table. */ | |
491 struct temp_slot_address_entry GTY(()) | |
492 { | |
493 hashval_t hash; | |
494 rtx address; | |
495 struct temp_slot *temp_slot; | |
496 }; | |
497 | |
498 /* Removes temporary slot TEMP from LIST. */ | |
499 | |
500 static void | |
501 cut_slot_from_list (struct temp_slot *temp, struct temp_slot **list) | |
502 { | |
503 if (temp->next) | |
504 temp->next->prev = temp->prev; | |
505 if (temp->prev) | |
506 temp->prev->next = temp->next; | |
507 else | |
508 *list = temp->next; | |
509 | |
510 temp->prev = temp->next = NULL; | |
511 } | |
512 | |
513 /* Inserts temporary slot TEMP to LIST. */ | |
514 | |
515 static void | |
516 insert_slot_to_list (struct temp_slot *temp, struct temp_slot **list) | |
517 { | |
518 temp->next = *list; | |
519 if (*list) | |
520 (*list)->prev = temp; | |
521 temp->prev = NULL; | |
522 *list = temp; | |
523 } | |
524 | |
525 /* Returns the list of used temp slots at LEVEL. */ | |
526 | |
527 static struct temp_slot ** | |
528 temp_slots_at_level (int level) | |
529 { | |
530 if (level >= (int) VEC_length (temp_slot_p, used_temp_slots)) | |
531 VEC_safe_grow_cleared (temp_slot_p, gc, used_temp_slots, level + 1); | |
532 | |
533 return &(VEC_address (temp_slot_p, used_temp_slots)[level]); | |
534 } | |
535 | |
536 /* Returns the maximal temporary slot level. */ | |
537 | |
538 static int | |
539 max_slot_level (void) | |
540 { | |
541 if (!used_temp_slots) | |
542 return -1; | |
543 | |
544 return VEC_length (temp_slot_p, used_temp_slots) - 1; | |
545 } | |
546 | |
547 /* Moves temporary slot TEMP to LEVEL. */ | |
548 | |
549 static void | |
550 move_slot_to_level (struct temp_slot *temp, int level) | |
551 { | |
552 cut_slot_from_list (temp, temp_slots_at_level (temp->level)); | |
553 insert_slot_to_list (temp, temp_slots_at_level (level)); | |
554 temp->level = level; | |
555 } | |
556 | |
557 /* Make temporary slot TEMP available. */ | |
558 | |
559 static void | |
560 make_slot_available (struct temp_slot *temp) | |
561 { | |
562 cut_slot_from_list (temp, temp_slots_at_level (temp->level)); | |
563 insert_slot_to_list (temp, &avail_temp_slots); | |
564 temp->in_use = 0; | |
565 temp->level = -1; | |
566 } | |
567 | |
568 /* Compute the hash value for an address -> temp slot mapping. | |
569 The value is cached on the mapping entry. */ | |
570 static hashval_t | |
571 temp_slot_address_compute_hash (struct temp_slot_address_entry *t) | |
572 { | |
573 int do_not_record = 0; | |
574 return hash_rtx (t->address, GET_MODE (t->address), | |
575 &do_not_record, NULL, false); | |
576 } | |
577 | |
578 /* Return the hash value for an address -> temp slot mapping. */ | |
579 static hashval_t | |
580 temp_slot_address_hash (const void *p) | |
581 { | |
582 const struct temp_slot_address_entry *t; | |
583 t = (const struct temp_slot_address_entry *) p; | |
584 return t->hash; | |
585 } | |
586 | |
587 /* Compare two address -> temp slot mapping entries. */ | |
588 static int | |
589 temp_slot_address_eq (const void *p1, const void *p2) | |
590 { | |
591 const struct temp_slot_address_entry *t1, *t2; | |
592 t1 = (const struct temp_slot_address_entry *) p1; | |
593 t2 = (const struct temp_slot_address_entry *) p2; | |
594 return exp_equiv_p (t1->address, t2->address, 0, true); | |
595 } | |
596 | |
597 /* Add ADDRESS as an alias of TEMP_SLOT to the addess -> temp slot mapping. */ | |
598 static void | |
599 insert_temp_slot_address (rtx address, struct temp_slot *temp_slot) | |
600 { | |
601 void **slot; | |
602 struct temp_slot_address_entry *t = GGC_NEW (struct temp_slot_address_entry); | |
603 t->address = address; | |
604 t->temp_slot = temp_slot; | |
605 t->hash = temp_slot_address_compute_hash (t); | |
606 slot = htab_find_slot_with_hash (temp_slot_address_table, t, t->hash, INSERT); | |
607 *slot = t; | |
608 } | |
609 | |
610 /* Remove an address -> temp slot mapping entry if the temp slot is | |
611 not in use anymore. Callback for remove_unused_temp_slot_addresses. */ | |
612 static int | |
613 remove_unused_temp_slot_addresses_1 (void **slot, void *data ATTRIBUTE_UNUSED) | |
614 { | |
615 const struct temp_slot_address_entry *t; | |
616 t = (const struct temp_slot_address_entry *) *slot; | |
617 if (! t->temp_slot->in_use) | |
618 *slot = NULL; | |
619 return 1; | |
620 } | |
621 | |
622 /* Remove all mappings of addresses to unused temp slots. */ | |
623 static void | |
624 remove_unused_temp_slot_addresses (void) | |
625 { | |
626 htab_traverse (temp_slot_address_table, | |
627 remove_unused_temp_slot_addresses_1, | |
628 NULL); | |
629 } | |
630 | |
631 /* Find the temp slot corresponding to the object at address X. */ | |
632 | |
633 static struct temp_slot * | |
634 find_temp_slot_from_address (rtx x) | |
635 { | |
636 struct temp_slot *p; | |
637 struct temp_slot_address_entry tmp, *t; | |
638 | |
639 /* First try the easy way: | |
640 See if X exists in the address -> temp slot mapping. */ | |
641 tmp.address = x; | |
642 tmp.temp_slot = NULL; | |
643 tmp.hash = temp_slot_address_compute_hash (&tmp); | |
644 t = (struct temp_slot_address_entry *) | |
645 htab_find_with_hash (temp_slot_address_table, &tmp, tmp.hash); | |
646 if (t) | |
647 return t->temp_slot; | |
648 | |
649 /* If we have a sum involving a register, see if it points to a temp | |
650 slot. */ | |
651 if (GET_CODE (x) == PLUS && REG_P (XEXP (x, 0)) | |
652 && (p = find_temp_slot_from_address (XEXP (x, 0))) != 0) | |
653 return p; | |
654 else if (GET_CODE (x) == PLUS && REG_P (XEXP (x, 1)) | |
655 && (p = find_temp_slot_from_address (XEXP (x, 1))) != 0) | |
656 return p; | |
657 | |
658 /* Last resort: Address is a virtual stack var address. */ | |
659 if (GET_CODE (x) == PLUS | |
660 && XEXP (x, 0) == virtual_stack_vars_rtx | |
661 && GET_CODE (XEXP (x, 1)) == CONST_INT) | |
662 { | |
663 int i; | |
664 for (i = max_slot_level (); i >= 0; i--) | |
665 for (p = *temp_slots_at_level (i); p; p = p->next) | |
666 { | |
667 if (INTVAL (XEXP (x, 1)) >= p->base_offset | |
668 && INTVAL (XEXP (x, 1)) < p->base_offset + p->full_size) | |
669 return p; | |
670 } | |
671 } | |
672 | |
673 return NULL; | |
674 } | |
675 | |
676 /* Allocate a temporary stack slot and record it for possible later | |
677 reuse. | |
678 | |
679 MODE is the machine mode to be given to the returned rtx. | |
680 | |
681 SIZE is the size in units of the space required. We do no rounding here | |
682 since assign_stack_local will do any required rounding. | |
683 | |
684 KEEP is 1 if this slot is to be retained after a call to | |
685 free_temp_slots. Automatic variables for a block are allocated | |
686 with this flag. KEEP values of 2 or 3 were needed respectively | |
687 for variables whose lifetime is controlled by CLEANUP_POINT_EXPRs | |
688 or for SAVE_EXPRs, but they are now unused. | |
689 | |
690 TYPE is the type that will be used for the stack slot. */ | |
691 | |
692 rtx | |
693 assign_stack_temp_for_type (enum machine_mode mode, HOST_WIDE_INT size, | |
694 int keep, tree type) | |
695 { | |
696 unsigned int align; | |
697 struct temp_slot *p, *best_p = 0, *selected = NULL, **pp; | |
698 rtx slot; | |
699 | |
700 /* If SIZE is -1 it means that somebody tried to allocate a temporary | |
701 of a variable size. */ | |
702 gcc_assert (size != -1); | |
703 | |
704 /* These are now unused. */ | |
705 gcc_assert (keep <= 1); | |
706 | |
707 align = get_stack_local_alignment (type, mode); | |
708 | |
709 /* Try to find an available, already-allocated temporary of the proper | |
710 mode which meets the size and alignment requirements. Choose the | |
711 smallest one with the closest alignment. | |
712 | |
713 If assign_stack_temp is called outside of the tree->rtl expansion, | |
714 we cannot reuse the stack slots (that may still refer to | |
715 VIRTUAL_STACK_VARS_REGNUM). */ | |
716 if (!virtuals_instantiated) | |
717 { | |
718 for (p = avail_temp_slots; p; p = p->next) | |
719 { | |
720 if (p->align >= align && p->size >= size | |
721 && GET_MODE (p->slot) == mode | |
722 && objects_must_conflict_p (p->type, type) | |
723 && (best_p == 0 || best_p->size > p->size | |
724 || (best_p->size == p->size && best_p->align > p->align))) | |
725 { | |
726 if (p->align == align && p->size == size) | |
727 { | |
728 selected = p; | |
729 cut_slot_from_list (selected, &avail_temp_slots); | |
730 best_p = 0; | |
731 break; | |
732 } | |
733 best_p = p; | |
734 } | |
735 } | |
736 } | |
737 | |
738 /* Make our best, if any, the one to use. */ | |
739 if (best_p) | |
740 { | |
741 selected = best_p; | |
742 cut_slot_from_list (selected, &avail_temp_slots); | |
743 | |
744 /* If there are enough aligned bytes left over, make them into a new | |
745 temp_slot so that the extra bytes don't get wasted. Do this only | |
746 for BLKmode slots, so that we can be sure of the alignment. */ | |
747 if (GET_MODE (best_p->slot) == BLKmode) | |
748 { | |
749 int alignment = best_p->align / BITS_PER_UNIT; | |
750 HOST_WIDE_INT rounded_size = CEIL_ROUND (size, alignment); | |
751 | |
752 if (best_p->size - rounded_size >= alignment) | |
753 { | |
754 p = GGC_NEW (struct temp_slot); | |
755 p->in_use = p->addr_taken = 0; | |
756 p->size = best_p->size - rounded_size; | |
757 p->base_offset = best_p->base_offset + rounded_size; | |
758 p->full_size = best_p->full_size - rounded_size; | |
759 p->slot = adjust_address_nv (best_p->slot, BLKmode, rounded_size); | |
760 p->align = best_p->align; | |
761 p->type = best_p->type; | |
762 insert_slot_to_list (p, &avail_temp_slots); | |
763 | |
764 stack_slot_list = gen_rtx_EXPR_LIST (VOIDmode, p->slot, | |
765 stack_slot_list); | |
766 | |
767 best_p->size = rounded_size; | |
768 best_p->full_size = rounded_size; | |
769 } | |
770 } | |
771 } | |
772 | |
773 /* If we still didn't find one, make a new temporary. */ | |
774 if (selected == 0) | |
775 { | |
776 HOST_WIDE_INT frame_offset_old = frame_offset; | |
777 | |
778 p = GGC_NEW (struct temp_slot); | |
779 | |
780 /* We are passing an explicit alignment request to assign_stack_local. | |
781 One side effect of that is assign_stack_local will not round SIZE | |
782 to ensure the frame offset remains suitably aligned. | |
783 | |
784 So for requests which depended on the rounding of SIZE, we go ahead | |
785 and round it now. We also make sure ALIGNMENT is at least | |
786 BIGGEST_ALIGNMENT. */ | |
787 gcc_assert (mode != BLKmode || align == BIGGEST_ALIGNMENT); | |
788 p->slot = assign_stack_local (mode, | |
789 (mode == BLKmode | |
790 ? CEIL_ROUND (size, (int) align / BITS_PER_UNIT) | |
791 : size), | |
792 align); | |
793 | |
794 p->align = align; | |
795 | |
796 /* The following slot size computation is necessary because we don't | |
797 know the actual size of the temporary slot until assign_stack_local | |
798 has performed all the frame alignment and size rounding for the | |
799 requested temporary. Note that extra space added for alignment | |
800 can be either above or below this stack slot depending on which | |
801 way the frame grows. We include the extra space if and only if it | |
802 is above this slot. */ | |
803 if (FRAME_GROWS_DOWNWARD) | |
804 p->size = frame_offset_old - frame_offset; | |
805 else | |
806 p->size = size; | |
807 | |
808 /* Now define the fields used by combine_temp_slots. */ | |
809 if (FRAME_GROWS_DOWNWARD) | |
810 { | |
811 p->base_offset = frame_offset; | |
812 p->full_size = frame_offset_old - frame_offset; | |
813 } | |
814 else | |
815 { | |
816 p->base_offset = frame_offset_old; | |
817 p->full_size = frame_offset - frame_offset_old; | |
818 } | |
819 | |
820 selected = p; | |
821 } | |
822 | |
823 p = selected; | |
824 p->in_use = 1; | |
825 p->addr_taken = 0; | |
826 p->type = type; | |
827 p->level = temp_slot_level; | |
828 p->keep = keep; | |
829 | |
830 pp = temp_slots_at_level (p->level); | |
831 insert_slot_to_list (p, pp); | |
832 insert_temp_slot_address (XEXP (p->slot, 0), p); | |
833 | |
834 /* Create a new MEM rtx to avoid clobbering MEM flags of old slots. */ | |
835 slot = gen_rtx_MEM (mode, XEXP (p->slot, 0)); | |
836 stack_slot_list = gen_rtx_EXPR_LIST (VOIDmode, slot, stack_slot_list); | |
837 | |
838 /* If we know the alias set for the memory that will be used, use | |
839 it. If there's no TYPE, then we don't know anything about the | |
840 alias set for the memory. */ | |
841 set_mem_alias_set (slot, type ? get_alias_set (type) : 0); | |
842 set_mem_align (slot, align); | |
843 | |
844 /* If a type is specified, set the relevant flags. */ | |
845 if (type != 0) | |
846 { | |
847 MEM_VOLATILE_P (slot) = TYPE_VOLATILE (type); | |
848 MEM_SET_IN_STRUCT_P (slot, (AGGREGATE_TYPE_P (type) | |
849 || TREE_CODE (type) == COMPLEX_TYPE)); | |
850 } | |
851 MEM_NOTRAP_P (slot) = 1; | |
852 | |
853 return slot; | |
854 } | |
855 | |
856 /* Allocate a temporary stack slot and record it for possible later | |
857 reuse. First three arguments are same as in preceding function. */ | |
858 | |
859 rtx | |
860 assign_stack_temp (enum machine_mode mode, HOST_WIDE_INT size, int keep) | |
861 { | |
862 return assign_stack_temp_for_type (mode, size, keep, NULL_TREE); | |
863 } | |
864 | |
865 /* Assign a temporary. | |
866 If TYPE_OR_DECL is a decl, then we are doing it on behalf of the decl | |
867 and so that should be used in error messages. In either case, we | |
868 allocate of the given type. | |
869 KEEP is as for assign_stack_temp. | |
870 MEMORY_REQUIRED is 1 if the result must be addressable stack memory; | |
871 it is 0 if a register is OK. | |
872 DONT_PROMOTE is 1 if we should not promote values in register | |
873 to wider modes. */ | |
874 | |
875 rtx | |
876 assign_temp (tree type_or_decl, int keep, int memory_required, | |
877 int dont_promote ATTRIBUTE_UNUSED) | |
878 { | |
879 tree type, decl; | |
880 enum machine_mode mode; | |
881 #ifdef PROMOTE_MODE | |
882 int unsignedp; | |
883 #endif | |
884 | |
885 if (DECL_P (type_or_decl)) | |
886 decl = type_or_decl, type = TREE_TYPE (decl); | |
887 else | |
888 decl = NULL, type = type_or_decl; | |
889 | |
890 mode = TYPE_MODE (type); | |
891 #ifdef PROMOTE_MODE | |
892 unsignedp = TYPE_UNSIGNED (type); | |
893 #endif | |
894 | |
895 if (mode == BLKmode || memory_required) | |
896 { | |
897 HOST_WIDE_INT size = int_size_in_bytes (type); | |
898 rtx tmp; | |
899 | |
900 /* Zero sized arrays are GNU C extension. Set size to 1 to avoid | |
901 problems with allocating the stack space. */ | |
902 if (size == 0) | |
903 size = 1; | |
904 | |
905 /* Unfortunately, we don't yet know how to allocate variable-sized | |
906 temporaries. However, sometimes we can find a fixed upper limit on | |
907 the size, so try that instead. */ | |
908 else if (size == -1) | |
909 size = max_int_size_in_bytes (type); | |
910 | |
911 /* The size of the temporary may be too large to fit into an integer. */ | |
912 /* ??? Not sure this should happen except for user silliness, so limit | |
913 this to things that aren't compiler-generated temporaries. The | |
914 rest of the time we'll die in assign_stack_temp_for_type. */ | |
915 if (decl && size == -1 | |
916 && TREE_CODE (TYPE_SIZE_UNIT (type)) == INTEGER_CST) | |
917 { | |
918 error ("size of variable %q+D is too large", decl); | |
919 size = 1; | |
920 } | |
921 | |
922 tmp = assign_stack_temp_for_type (mode, size, keep, type); | |
923 return tmp; | |
924 } | |
925 | |
926 #ifdef PROMOTE_MODE | |
927 if (! dont_promote) | |
928 mode = promote_mode (type, mode, &unsignedp, 0); | |
929 #endif | |
930 | |
931 return gen_reg_rtx (mode); | |
932 } | |
933 | |
934 /* Combine temporary stack slots which are adjacent on the stack. | |
935 | |
936 This allows for better use of already allocated stack space. This is only | |
937 done for BLKmode slots because we can be sure that we won't have alignment | |
938 problems in this case. */ | |
939 | |
940 static void | |
941 combine_temp_slots (void) | |
942 { | |
943 struct temp_slot *p, *q, *next, *next_q; | |
944 int num_slots; | |
945 | |
946 /* We can't combine slots, because the information about which slot | |
947 is in which alias set will be lost. */ | |
948 if (flag_strict_aliasing) | |
949 return; | |
950 | |
951 /* If there are a lot of temp slots, don't do anything unless | |
952 high levels of optimization. */ | |
953 if (! flag_expensive_optimizations) | |
954 for (p = avail_temp_slots, num_slots = 0; p; p = p->next, num_slots++) | |
955 if (num_slots > 100 || (num_slots > 10 && optimize == 0)) | |
956 return; | |
957 | |
958 for (p = avail_temp_slots; p; p = next) | |
959 { | |
960 int delete_p = 0; | |
961 | |
962 next = p->next; | |
963 | |
964 if (GET_MODE (p->slot) != BLKmode) | |
965 continue; | |
966 | |
967 for (q = p->next; q; q = next_q) | |
968 { | |
969 int delete_q = 0; | |
970 | |
971 next_q = q->next; | |
972 | |
973 if (GET_MODE (q->slot) != BLKmode) | |
974 continue; | |
975 | |
976 if (p->base_offset + p->full_size == q->base_offset) | |
977 { | |
978 /* Q comes after P; combine Q into P. */ | |
979 p->size += q->size; | |
980 p->full_size += q->full_size; | |
981 delete_q = 1; | |
982 } | |
983 else if (q->base_offset + q->full_size == p->base_offset) | |
984 { | |
985 /* P comes after Q; combine P into Q. */ | |
986 q->size += p->size; | |
987 q->full_size += p->full_size; | |
988 delete_p = 1; | |
989 break; | |
990 } | |
991 if (delete_q) | |
992 cut_slot_from_list (q, &avail_temp_slots); | |
993 } | |
994 | |
995 /* Either delete P or advance past it. */ | |
996 if (delete_p) | |
997 cut_slot_from_list (p, &avail_temp_slots); | |
998 } | |
999 } | |
1000 | |
1001 /* Indicate that NEW_RTX is an alternate way of referring to the temp | |
1002 slot that previously was known by OLD_RTX. */ | |
1003 | |
1004 void | |
1005 update_temp_slot_address (rtx old_rtx, rtx new_rtx) | |
1006 { | |
1007 struct temp_slot *p; | |
1008 | |
1009 if (rtx_equal_p (old_rtx, new_rtx)) | |
1010 return; | |
1011 | |
1012 p = find_temp_slot_from_address (old_rtx); | |
1013 | |
1014 /* If we didn't find one, see if both OLD_RTX is a PLUS. If so, and | |
1015 NEW_RTX is a register, see if one operand of the PLUS is a | |
1016 temporary location. If so, NEW_RTX points into it. Otherwise, | |
1017 if both OLD_RTX and NEW_RTX are a PLUS and if there is a register | |
1018 in common between them. If so, try a recursive call on those | |
1019 values. */ | |
1020 if (p == 0) | |
1021 { | |
1022 if (GET_CODE (old_rtx) != PLUS) | |
1023 return; | |
1024 | |
1025 if (REG_P (new_rtx)) | |
1026 { | |
1027 update_temp_slot_address (XEXP (old_rtx, 0), new_rtx); | |
1028 update_temp_slot_address (XEXP (old_rtx, 1), new_rtx); | |
1029 return; | |
1030 } | |
1031 else if (GET_CODE (new_rtx) != PLUS) | |
1032 return; | |
1033 | |
1034 if (rtx_equal_p (XEXP (old_rtx, 0), XEXP (new_rtx, 0))) | |
1035 update_temp_slot_address (XEXP (old_rtx, 1), XEXP (new_rtx, 1)); | |
1036 else if (rtx_equal_p (XEXP (old_rtx, 1), XEXP (new_rtx, 0))) | |
1037 update_temp_slot_address (XEXP (old_rtx, 0), XEXP (new_rtx, 1)); | |
1038 else if (rtx_equal_p (XEXP (old_rtx, 0), XEXP (new_rtx, 1))) | |
1039 update_temp_slot_address (XEXP (old_rtx, 1), XEXP (new_rtx, 0)); | |
1040 else if (rtx_equal_p (XEXP (old_rtx, 1), XEXP (new_rtx, 1))) | |
1041 update_temp_slot_address (XEXP (old_rtx, 0), XEXP (new_rtx, 0)); | |
1042 | |
1043 return; | |
1044 } | |
1045 | |
1046 /* Otherwise add an alias for the temp's address. */ | |
1047 insert_temp_slot_address (new_rtx, p); | |
1048 } | |
1049 | |
1050 /* If X could be a reference to a temporary slot, mark the fact that its | |
1051 address was taken. */ | |
1052 | |
1053 void | |
1054 mark_temp_addr_taken (rtx x) | |
1055 { | |
1056 struct temp_slot *p; | |
1057 | |
1058 if (x == 0) | |
1059 return; | |
1060 | |
1061 /* If X is not in memory or is at a constant address, it cannot be in | |
1062 a temporary slot. */ | |
1063 if (!MEM_P (x) || CONSTANT_P (XEXP (x, 0))) | |
1064 return; | |
1065 | |
1066 p = find_temp_slot_from_address (XEXP (x, 0)); | |
1067 if (p != 0) | |
1068 p->addr_taken = 1; | |
1069 } | |
1070 | |
1071 /* If X could be a reference to a temporary slot, mark that slot as | |
1072 belonging to the to one level higher than the current level. If X | |
1073 matched one of our slots, just mark that one. Otherwise, we can't | |
1074 easily predict which it is, so upgrade all of them. Kept slots | |
1075 need not be touched. | |
1076 | |
1077 This is called when an ({...}) construct occurs and a statement | |
1078 returns a value in memory. */ | |
1079 | |
1080 void | |
1081 preserve_temp_slots (rtx x) | |
1082 { | |
1083 struct temp_slot *p = 0, *next; | |
1084 | |
1085 /* If there is no result, we still might have some objects whose address | |
1086 were taken, so we need to make sure they stay around. */ | |
1087 if (x == 0) | |
1088 { | |
1089 for (p = *temp_slots_at_level (temp_slot_level); p; p = next) | |
1090 { | |
1091 next = p->next; | |
1092 | |
1093 if (p->addr_taken) | |
1094 move_slot_to_level (p, temp_slot_level - 1); | |
1095 } | |
1096 | |
1097 return; | |
1098 } | |
1099 | |
1100 /* If X is a register that is being used as a pointer, see if we have | |
1101 a temporary slot we know it points to. To be consistent with | |
1102 the code below, we really should preserve all non-kept slots | |
1103 if we can't find a match, but that seems to be much too costly. */ | |
1104 if (REG_P (x) && REG_POINTER (x)) | |
1105 p = find_temp_slot_from_address (x); | |
1106 | |
1107 /* If X is not in memory or is at a constant address, it cannot be in | |
1108 a temporary slot, but it can contain something whose address was | |
1109 taken. */ | |
1110 if (p == 0 && (!MEM_P (x) || CONSTANT_P (XEXP (x, 0)))) | |
1111 { | |
1112 for (p = *temp_slots_at_level (temp_slot_level); p; p = next) | |
1113 { | |
1114 next = p->next; | |
1115 | |
1116 if (p->addr_taken) | |
1117 move_slot_to_level (p, temp_slot_level - 1); | |
1118 } | |
1119 | |
1120 return; | |
1121 } | |
1122 | |
1123 /* First see if we can find a match. */ | |
1124 if (p == 0) | |
1125 p = find_temp_slot_from_address (XEXP (x, 0)); | |
1126 | |
1127 if (p != 0) | |
1128 { | |
1129 /* Move everything at our level whose address was taken to our new | |
1130 level in case we used its address. */ | |
1131 struct temp_slot *q; | |
1132 | |
1133 if (p->level == temp_slot_level) | |
1134 { | |
1135 for (q = *temp_slots_at_level (temp_slot_level); q; q = next) | |
1136 { | |
1137 next = q->next; | |
1138 | |
1139 if (p != q && q->addr_taken) | |
1140 move_slot_to_level (q, temp_slot_level - 1); | |
1141 } | |
1142 | |
1143 move_slot_to_level (p, temp_slot_level - 1); | |
1144 p->addr_taken = 0; | |
1145 } | |
1146 return; | |
1147 } | |
1148 | |
1149 /* Otherwise, preserve all non-kept slots at this level. */ | |
1150 for (p = *temp_slots_at_level (temp_slot_level); p; p = next) | |
1151 { | |
1152 next = p->next; | |
1153 | |
1154 if (!p->keep) | |
1155 move_slot_to_level (p, temp_slot_level - 1); | |
1156 } | |
1157 } | |
1158 | |
1159 /* Free all temporaries used so far. This is normally called at the | |
1160 end of generating code for a statement. */ | |
1161 | |
1162 void | |
1163 free_temp_slots (void) | |
1164 { | |
1165 struct temp_slot *p, *next; | |
1166 | |
1167 for (p = *temp_slots_at_level (temp_slot_level); p; p = next) | |
1168 { | |
1169 next = p->next; | |
1170 | |
1171 if (!p->keep) | |
1172 make_slot_available (p); | |
1173 } | |
1174 | |
1175 remove_unused_temp_slot_addresses (); | |
1176 combine_temp_slots (); | |
1177 } | |
1178 | |
1179 /* Push deeper into the nesting level for stack temporaries. */ | |
1180 | |
1181 void | |
1182 push_temp_slots (void) | |
1183 { | |
1184 temp_slot_level++; | |
1185 } | |
1186 | |
1187 /* Pop a temporary nesting level. All slots in use in the current level | |
1188 are freed. */ | |
1189 | |
1190 void | |
1191 pop_temp_slots (void) | |
1192 { | |
1193 struct temp_slot *p, *next; | |
1194 | |
1195 for (p = *temp_slots_at_level (temp_slot_level); p; p = next) | |
1196 { | |
1197 next = p->next; | |
1198 make_slot_available (p); | |
1199 } | |
1200 | |
1201 remove_unused_temp_slot_addresses (); | |
1202 combine_temp_slots (); | |
1203 | |
1204 temp_slot_level--; | |
1205 } | |
1206 | |
1207 /* Initialize temporary slots. */ | |
1208 | |
1209 void | |
1210 init_temp_slots (void) | |
1211 { | |
1212 /* We have not allocated any temporaries yet. */ | |
1213 avail_temp_slots = 0; | |
1214 used_temp_slots = 0; | |
1215 temp_slot_level = 0; | |
1216 | |
1217 /* Set up the table to map addresses to temp slots. */ | |
1218 if (! temp_slot_address_table) | |
1219 temp_slot_address_table = htab_create_ggc (32, | |
1220 temp_slot_address_hash, | |
1221 temp_slot_address_eq, | |
1222 NULL); | |
1223 else | |
1224 htab_empty (temp_slot_address_table); | |
1225 } | |
1226 | |
1227 /* These routines are responsible for converting virtual register references | |
1228 to the actual hard register references once RTL generation is complete. | |
1229 | |
1230 The following four variables are used for communication between the | |
1231 routines. They contain the offsets of the virtual registers from their | |
1232 respective hard registers. */ | |
1233 | |
1234 static int in_arg_offset; | |
1235 static int var_offset; | |
1236 static int dynamic_offset; | |
1237 static int out_arg_offset; | |
1238 static int cfa_offset; | |
1239 | |
1240 /* In most machines, the stack pointer register is equivalent to the bottom | |
1241 of the stack. */ | |
1242 | |
1243 #ifndef STACK_POINTER_OFFSET | |
1244 #define STACK_POINTER_OFFSET 0 | |
1245 #endif | |
1246 | |
1247 /* If not defined, pick an appropriate default for the offset of dynamically | |
1248 allocated memory depending on the value of ACCUMULATE_OUTGOING_ARGS, | |
1249 REG_PARM_STACK_SPACE, and OUTGOING_REG_PARM_STACK_SPACE. */ | |
1250 | |
1251 #ifndef STACK_DYNAMIC_OFFSET | |
1252 | |
1253 /* The bottom of the stack points to the actual arguments. If | |
1254 REG_PARM_STACK_SPACE is defined, this includes the space for the register | |
1255 parameters. However, if OUTGOING_REG_PARM_STACK space is not defined, | |
1256 stack space for register parameters is not pushed by the caller, but | |
1257 rather part of the fixed stack areas and hence not included in | |
1258 `crtl->outgoing_args_size'. Nevertheless, we must allow | |
1259 for it when allocating stack dynamic objects. */ | |
1260 | |
1261 #if defined(REG_PARM_STACK_SPACE) | |
1262 #define STACK_DYNAMIC_OFFSET(FNDECL) \ | |
1263 ((ACCUMULATE_OUTGOING_ARGS \ | |
1264 ? (crtl->outgoing_args_size \ | |
1265 + (OUTGOING_REG_PARM_STACK_SPACE ((!(FNDECL) ? NULL_TREE : TREE_TYPE (FNDECL))) ? 0 \ | |
1266 : REG_PARM_STACK_SPACE (FNDECL))) \ | |
1267 : 0) + (STACK_POINTER_OFFSET)) | |
1268 #else | |
1269 #define STACK_DYNAMIC_OFFSET(FNDECL) \ | |
1270 ((ACCUMULATE_OUTGOING_ARGS ? crtl->outgoing_args_size : 0) \ | |
1271 + (STACK_POINTER_OFFSET)) | |
1272 #endif | |
1273 #endif | |
1274 | |
1275 | |
1276 /* Given a piece of RTX and a pointer to a HOST_WIDE_INT, if the RTX | |
1277 is a virtual register, return the equivalent hard register and set the | |
1278 offset indirectly through the pointer. Otherwise, return 0. */ | |
1279 | |
1280 static rtx | |
1281 instantiate_new_reg (rtx x, HOST_WIDE_INT *poffset) | |
1282 { | |
1283 rtx new_rtx; | |
1284 HOST_WIDE_INT offset; | |
1285 | |
1286 if (x == virtual_incoming_args_rtx) | |
1287 { | |
1288 if (stack_realign_drap) | |
1289 { | |
1290 /* Replace virtual_incoming_args_rtx with internal arg | |
1291 pointer if DRAP is used to realign stack. */ | |
1292 new_rtx = crtl->args.internal_arg_pointer; | |
1293 offset = 0; | |
1294 } | |
1295 else | |
1296 new_rtx = arg_pointer_rtx, offset = in_arg_offset; | |
1297 } | |
1298 else if (x == virtual_stack_vars_rtx) | |
1299 new_rtx = frame_pointer_rtx, offset = var_offset; | |
1300 else if (x == virtual_stack_dynamic_rtx) | |
1301 new_rtx = stack_pointer_rtx, offset = dynamic_offset; | |
1302 else if (x == virtual_outgoing_args_rtx) | |
1303 new_rtx = stack_pointer_rtx, offset = out_arg_offset; | |
1304 else if (x == virtual_cfa_rtx) | |
1305 { | |
1306 #ifdef FRAME_POINTER_CFA_OFFSET | |
1307 new_rtx = frame_pointer_rtx; | |
1308 #else | |
1309 new_rtx = arg_pointer_rtx; | |
1310 #endif | |
1311 offset = cfa_offset; | |
1312 } | |
1313 else | |
1314 return NULL_RTX; | |
1315 | |
1316 *poffset = offset; | |
1317 return new_rtx; | |
1318 } | |
1319 | |
1320 /* A subroutine of instantiate_virtual_regs, called via for_each_rtx. | |
1321 Instantiate any virtual registers present inside of *LOC. The expression | |
1322 is simplified, as much as possible, but is not to be considered "valid" | |
1323 in any sense implied by the target. If any change is made, set CHANGED | |
1324 to true. */ | |
1325 | |
1326 static int | |
1327 instantiate_virtual_regs_in_rtx (rtx *loc, void *data) | |
1328 { | |
1329 HOST_WIDE_INT offset; | |
1330 bool *changed = (bool *) data; | |
1331 rtx x, new_rtx; | |
1332 | |
1333 x = *loc; | |
1334 if (x == 0) | |
1335 return 0; | |
1336 | |
1337 switch (GET_CODE (x)) | |
1338 { | |
1339 case REG: | |
1340 new_rtx = instantiate_new_reg (x, &offset); | |
1341 if (new_rtx) | |
1342 { | |
1343 *loc = plus_constant (new_rtx, offset); | |
1344 if (changed) | |
1345 *changed = true; | |
1346 } | |
1347 return -1; | |
1348 | |
1349 case PLUS: | |
1350 new_rtx = instantiate_new_reg (XEXP (x, 0), &offset); | |
1351 if (new_rtx) | |
1352 { | |
1353 new_rtx = plus_constant (new_rtx, offset); | |
1354 *loc = simplify_gen_binary (PLUS, GET_MODE (x), new_rtx, XEXP (x, 1)); | |
1355 if (changed) | |
1356 *changed = true; | |
1357 return -1; | |
1358 } | |
1359 | |
1360 /* FIXME -- from old code */ | |
1361 /* If we have (plus (subreg (virtual-reg)) (const_int)), we know | |
1362 we can commute the PLUS and SUBREG because pointers into the | |
1363 frame are well-behaved. */ | |
1364 break; | |
1365 | |
1366 default: | |
1367 break; | |
1368 } | |
1369 | |
1370 return 0; | |
1371 } | |
1372 | |
1373 /* A subroutine of instantiate_virtual_regs_in_insn. Return true if X | |
1374 matches the predicate for insn CODE operand OPERAND. */ | |
1375 | |
1376 static int | |
1377 safe_insn_predicate (int code, int operand, rtx x) | |
1378 { | |
1379 const struct insn_operand_data *op_data; | |
1380 | |
1381 if (code < 0) | |
1382 return true; | |
1383 | |
1384 op_data = &insn_data[code].operand[operand]; | |
1385 if (op_data->predicate == NULL) | |
1386 return true; | |
1387 | |
1388 return op_data->predicate (x, op_data->mode); | |
1389 } | |
1390 | |
1391 /* A subroutine of instantiate_virtual_regs. Instantiate any virtual | |
1392 registers present inside of insn. The result will be a valid insn. */ | |
1393 | |
1394 static void | |
1395 instantiate_virtual_regs_in_insn (rtx insn) | |
1396 { | |
1397 HOST_WIDE_INT offset; | |
1398 int insn_code, i; | |
1399 bool any_change = false; | |
1400 rtx set, new_rtx, x, seq; | |
1401 | |
1402 /* There are some special cases to be handled first. */ | |
1403 set = single_set (insn); | |
1404 if (set) | |
1405 { | |
1406 /* We're allowed to assign to a virtual register. This is interpreted | |
1407 to mean that the underlying register gets assigned the inverse | |
1408 transformation. This is used, for example, in the handling of | |
1409 non-local gotos. */ | |
1410 new_rtx = instantiate_new_reg (SET_DEST (set), &offset); | |
1411 if (new_rtx) | |
1412 { | |
1413 start_sequence (); | |
1414 | |
1415 for_each_rtx (&SET_SRC (set), instantiate_virtual_regs_in_rtx, NULL); | |
1416 x = simplify_gen_binary (PLUS, GET_MODE (new_rtx), SET_SRC (set), | |
1417 GEN_INT (-offset)); | |
1418 x = force_operand (x, new_rtx); | |
1419 if (x != new_rtx) | |
1420 emit_move_insn (new_rtx, x); | |
1421 | |
1422 seq = get_insns (); | |
1423 end_sequence (); | |
1424 | |
1425 emit_insn_before (seq, insn); | |
1426 delete_insn (insn); | |
1427 return; | |
1428 } | |
1429 | |
1430 /* Handle a straight copy from a virtual register by generating a | |
1431 new add insn. The difference between this and falling through | |
1432 to the generic case is avoiding a new pseudo and eliminating a | |
1433 move insn in the initial rtl stream. */ | |
1434 new_rtx = instantiate_new_reg (SET_SRC (set), &offset); | |
1435 if (new_rtx && offset != 0 | |
1436 && REG_P (SET_DEST (set)) | |
1437 && REGNO (SET_DEST (set)) > LAST_VIRTUAL_REGISTER) | |
1438 { | |
1439 start_sequence (); | |
1440 | |
1441 x = expand_simple_binop (GET_MODE (SET_DEST (set)), PLUS, | |
1442 new_rtx, GEN_INT (offset), SET_DEST (set), | |
1443 1, OPTAB_LIB_WIDEN); | |
1444 if (x != SET_DEST (set)) | |
1445 emit_move_insn (SET_DEST (set), x); | |
1446 | |
1447 seq = get_insns (); | |
1448 end_sequence (); | |
1449 | |
1450 emit_insn_before (seq, insn); | |
1451 delete_insn (insn); | |
1452 return; | |
1453 } | |
1454 | |
1455 extract_insn (insn); | |
1456 insn_code = INSN_CODE (insn); | |
1457 | |
1458 /* Handle a plus involving a virtual register by determining if the | |
1459 operands remain valid if they're modified in place. */ | |
1460 if (GET_CODE (SET_SRC (set)) == PLUS | |
1461 && recog_data.n_operands >= 3 | |
1462 && recog_data.operand_loc[1] == &XEXP (SET_SRC (set), 0) | |
1463 && recog_data.operand_loc[2] == &XEXP (SET_SRC (set), 1) | |
1464 && GET_CODE (recog_data.operand[2]) == CONST_INT | |
1465 && (new_rtx = instantiate_new_reg (recog_data.operand[1], &offset))) | |
1466 { | |
1467 offset += INTVAL (recog_data.operand[2]); | |
1468 | |
1469 /* If the sum is zero, then replace with a plain move. */ | |
1470 if (offset == 0 | |
1471 && REG_P (SET_DEST (set)) | |
1472 && REGNO (SET_DEST (set)) > LAST_VIRTUAL_REGISTER) | |
1473 { | |
1474 start_sequence (); | |
1475 emit_move_insn (SET_DEST (set), new_rtx); | |
1476 seq = get_insns (); | |
1477 end_sequence (); | |
1478 | |
1479 emit_insn_before (seq, insn); | |
1480 delete_insn (insn); | |
1481 return; | |
1482 } | |
1483 | |
1484 x = gen_int_mode (offset, recog_data.operand_mode[2]); | |
1485 | |
1486 /* Using validate_change and apply_change_group here leaves | |
1487 recog_data in an invalid state. Since we know exactly what | |
1488 we want to check, do those two by hand. */ | |
1489 if (safe_insn_predicate (insn_code, 1, new_rtx) | |
1490 && safe_insn_predicate (insn_code, 2, x)) | |
1491 { | |
1492 *recog_data.operand_loc[1] = recog_data.operand[1] = new_rtx; | |
1493 *recog_data.operand_loc[2] = recog_data.operand[2] = x; | |
1494 any_change = true; | |
1495 | |
1496 /* Fall through into the regular operand fixup loop in | |
1497 order to take care of operands other than 1 and 2. */ | |
1498 } | |
1499 } | |
1500 } | |
1501 else | |
1502 { | |
1503 extract_insn (insn); | |
1504 insn_code = INSN_CODE (insn); | |
1505 } | |
1506 | |
1507 /* In the general case, we expect virtual registers to appear only in | |
1508 operands, and then only as either bare registers or inside memories. */ | |
1509 for (i = 0; i < recog_data.n_operands; ++i) | |
1510 { | |
1511 x = recog_data.operand[i]; | |
1512 switch (GET_CODE (x)) | |
1513 { | |
1514 case MEM: | |
1515 { | |
1516 rtx addr = XEXP (x, 0); | |
1517 bool changed = false; | |
1518 | |
1519 for_each_rtx (&addr, instantiate_virtual_regs_in_rtx, &changed); | |
1520 if (!changed) | |
1521 continue; | |
1522 | |
1523 start_sequence (); | |
1524 x = replace_equiv_address (x, addr); | |
1525 /* It may happen that the address with the virtual reg | |
1526 was valid (e.g. based on the virtual stack reg, which might | |
1527 be acceptable to the predicates with all offsets), whereas | |
1528 the address now isn't anymore, for instance when the address | |
1529 is still offsetted, but the base reg isn't virtual-stack-reg | |
1530 anymore. Below we would do a force_reg on the whole operand, | |
1531 but this insn might actually only accept memory. Hence, | |
1532 before doing that last resort, try to reload the address into | |
1533 a register, so this operand stays a MEM. */ | |
1534 if (!safe_insn_predicate (insn_code, i, x)) | |
1535 { | |
1536 addr = force_reg (GET_MODE (addr), addr); | |
1537 x = replace_equiv_address (x, addr); | |
1538 } | |
1539 seq = get_insns (); | |
1540 end_sequence (); | |
1541 if (seq) | |
1542 emit_insn_before (seq, insn); | |
1543 } | |
1544 break; | |
1545 | |
1546 case REG: | |
1547 new_rtx = instantiate_new_reg (x, &offset); | |
1548 if (new_rtx == NULL) | |
1549 continue; | |
1550 if (offset == 0) | |
1551 x = new_rtx; | |
1552 else | |
1553 { | |
1554 start_sequence (); | |
1555 | |
1556 /* Careful, special mode predicates may have stuff in | |
1557 insn_data[insn_code].operand[i].mode that isn't useful | |
1558 to us for computing a new value. */ | |
1559 /* ??? Recognize address_operand and/or "p" constraints | |
1560 to see if (plus new offset) is a valid before we put | |
1561 this through expand_simple_binop. */ | |
1562 x = expand_simple_binop (GET_MODE (x), PLUS, new_rtx, | |
1563 GEN_INT (offset), NULL_RTX, | |
1564 1, OPTAB_LIB_WIDEN); | |
1565 seq = get_insns (); | |
1566 end_sequence (); | |
1567 emit_insn_before (seq, insn); | |
1568 } | |
1569 break; | |
1570 | |
1571 case SUBREG: | |
1572 new_rtx = instantiate_new_reg (SUBREG_REG (x), &offset); | |
1573 if (new_rtx == NULL) | |
1574 continue; | |
1575 if (offset != 0) | |
1576 { | |
1577 start_sequence (); | |
1578 new_rtx = expand_simple_binop (GET_MODE (new_rtx), PLUS, new_rtx, | |
1579 GEN_INT (offset), NULL_RTX, | |
1580 1, OPTAB_LIB_WIDEN); | |
1581 seq = get_insns (); | |
1582 end_sequence (); | |
1583 emit_insn_before (seq, insn); | |
1584 } | |
1585 x = simplify_gen_subreg (recog_data.operand_mode[i], new_rtx, | |
1586 GET_MODE (new_rtx), SUBREG_BYTE (x)); | |
1587 gcc_assert (x); | |
1588 break; | |
1589 | |
1590 default: | |
1591 continue; | |
1592 } | |
1593 | |
1594 /* At this point, X contains the new value for the operand. | |
1595 Validate the new value vs the insn predicate. Note that | |
1596 asm insns will have insn_code -1 here. */ | |
1597 if (!safe_insn_predicate (insn_code, i, x)) | |
1598 { | |
1599 start_sequence (); | |
1600 x = force_reg (insn_data[insn_code].operand[i].mode, x); | |
1601 seq = get_insns (); | |
1602 end_sequence (); | |
1603 if (seq) | |
1604 emit_insn_before (seq, insn); | |
1605 } | |
1606 | |
1607 *recog_data.operand_loc[i] = recog_data.operand[i] = x; | |
1608 any_change = true; | |
1609 } | |
1610 | |
1611 if (any_change) | |
1612 { | |
1613 /* Propagate operand changes into the duplicates. */ | |
1614 for (i = 0; i < recog_data.n_dups; ++i) | |
1615 *recog_data.dup_loc[i] | |
1616 = copy_rtx (recog_data.operand[(unsigned)recog_data.dup_num[i]]); | |
1617 | |
1618 /* Force re-recognition of the instruction for validation. */ | |
1619 INSN_CODE (insn) = -1; | |
1620 } | |
1621 | |
1622 if (asm_noperands (PATTERN (insn)) >= 0) | |
1623 { | |
1624 if (!check_asm_operands (PATTERN (insn))) | |
1625 { | |
1626 error_for_asm (insn, "impossible constraint in %<asm%>"); | |
1627 delete_insn (insn); | |
1628 } | |
1629 } | |
1630 else | |
1631 { | |
1632 if (recog_memoized (insn) < 0) | |
1633 fatal_insn_not_found (insn); | |
1634 } | |
1635 } | |
1636 | |
1637 /* Subroutine of instantiate_decls. Given RTL representing a decl, | |
1638 do any instantiation required. */ | |
1639 | |
1640 void | |
1641 instantiate_decl_rtl (rtx x) | |
1642 { | |
1643 rtx addr; | |
1644 | |
1645 if (x == 0) | |
1646 return; | |
1647 | |
1648 /* If this is a CONCAT, recurse for the pieces. */ | |
1649 if (GET_CODE (x) == CONCAT) | |
1650 { | |
1651 instantiate_decl_rtl (XEXP (x, 0)); | |
1652 instantiate_decl_rtl (XEXP (x, 1)); | |
1653 return; | |
1654 } | |
1655 | |
1656 /* If this is not a MEM, no need to do anything. Similarly if the | |
1657 address is a constant or a register that is not a virtual register. */ | |
1658 if (!MEM_P (x)) | |
1659 return; | |
1660 | |
1661 addr = XEXP (x, 0); | |
1662 if (CONSTANT_P (addr) | |
1663 || (REG_P (addr) | |
1664 && (REGNO (addr) < FIRST_VIRTUAL_REGISTER | |
1665 || REGNO (addr) > LAST_VIRTUAL_REGISTER))) | |
1666 return; | |
1667 | |
1668 for_each_rtx (&XEXP (x, 0), instantiate_virtual_regs_in_rtx, NULL); | |
1669 } | |
1670 | |
1671 /* Helper for instantiate_decls called via walk_tree: Process all decls | |
1672 in the given DECL_VALUE_EXPR. */ | |
1673 | |
1674 static tree | |
1675 instantiate_expr (tree *tp, int *walk_subtrees, void *data ATTRIBUTE_UNUSED) | |
1676 { | |
1677 tree t = *tp; | |
1678 if (! EXPR_P (t)) | |
1679 { | |
1680 *walk_subtrees = 0; | |
1681 if (DECL_P (t) && DECL_RTL_SET_P (t)) | |
1682 instantiate_decl_rtl (DECL_RTL (t)); | |
1683 } | |
1684 return NULL; | |
1685 } | |
1686 | |
1687 /* Subroutine of instantiate_decls: Process all decls in the given | |
1688 BLOCK node and all its subblocks. */ | |
1689 | |
1690 static void | |
1691 instantiate_decls_1 (tree let) | |
1692 { | |
1693 tree t; | |
1694 | |
1695 for (t = BLOCK_VARS (let); t; t = TREE_CHAIN (t)) | |
1696 { | |
1697 if (DECL_RTL_SET_P (t)) | |
1698 instantiate_decl_rtl (DECL_RTL (t)); | |
1699 if (TREE_CODE (t) == VAR_DECL && DECL_HAS_VALUE_EXPR_P (t)) | |
1700 { | |
1701 tree v = DECL_VALUE_EXPR (t); | |
1702 walk_tree (&v, instantiate_expr, NULL, NULL); | |
1703 } | |
1704 } | |
1705 | |
1706 /* Process all subblocks. */ | |
1707 for (t = BLOCK_SUBBLOCKS (let); t; t = BLOCK_CHAIN (t)) | |
1708 instantiate_decls_1 (t); | |
1709 } | |
1710 | |
1711 /* Scan all decls in FNDECL (both variables and parameters) and instantiate | |
1712 all virtual registers in their DECL_RTL's. */ | |
1713 | |
1714 static void | |
1715 instantiate_decls (tree fndecl) | |
1716 { | |
1717 tree decl, t, next; | |
1718 | |
1719 /* Process all parameters of the function. */ | |
1720 for (decl = DECL_ARGUMENTS (fndecl); decl; decl = TREE_CHAIN (decl)) | |
1721 { | |
1722 instantiate_decl_rtl (DECL_RTL (decl)); | |
1723 instantiate_decl_rtl (DECL_INCOMING_RTL (decl)); | |
1724 if (DECL_HAS_VALUE_EXPR_P (decl)) | |
1725 { | |
1726 tree v = DECL_VALUE_EXPR (decl); | |
1727 walk_tree (&v, instantiate_expr, NULL, NULL); | |
1728 } | |
1729 } | |
1730 | |
1731 /* Now process all variables defined in the function or its subblocks. */ | |
1732 instantiate_decls_1 (DECL_INITIAL (fndecl)); | |
1733 | |
1734 t = cfun->local_decls; | |
1735 cfun->local_decls = NULL_TREE; | |
1736 for (; t; t = next) | |
1737 { | |
1738 next = TREE_CHAIN (t); | |
1739 decl = TREE_VALUE (t); | |
1740 if (DECL_RTL_SET_P (decl)) | |
1741 instantiate_decl_rtl (DECL_RTL (decl)); | |
1742 ggc_free (t); | |
1743 } | |
1744 } | |
1745 | |
1746 /* Pass through the INSNS of function FNDECL and convert virtual register | |
1747 references to hard register references. */ | |
1748 | |
1749 static unsigned int | |
1750 instantiate_virtual_regs (void) | |
1751 { | |
1752 rtx insn; | |
1753 | |
1754 /* Compute the offsets to use for this function. */ | |
1755 in_arg_offset = FIRST_PARM_OFFSET (current_function_decl); | |
1756 var_offset = STARTING_FRAME_OFFSET; | |
1757 dynamic_offset = STACK_DYNAMIC_OFFSET (current_function_decl); | |
1758 out_arg_offset = STACK_POINTER_OFFSET; | |
1759 #ifdef FRAME_POINTER_CFA_OFFSET | |
1760 cfa_offset = FRAME_POINTER_CFA_OFFSET (current_function_decl); | |
1761 #else | |
1762 cfa_offset = ARG_POINTER_CFA_OFFSET (current_function_decl); | |
1763 #endif | |
1764 | |
1765 /* Initialize recognition, indicating that volatile is OK. */ | |
1766 init_recog (); | |
1767 | |
1768 /* Scan through all the insns, instantiating every virtual register still | |
1769 present. */ | |
1770 for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
1771 if (INSN_P (insn)) | |
1772 { | |
1773 /* These patterns in the instruction stream can never be recognized. | |
1774 Fortunately, they shouldn't contain virtual registers either. */ | |
1775 if (GET_CODE (PATTERN (insn)) == USE | |
1776 || GET_CODE (PATTERN (insn)) == CLOBBER | |
1777 || GET_CODE (PATTERN (insn)) == ADDR_VEC | |
1778 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC | |
1779 || GET_CODE (PATTERN (insn)) == ASM_INPUT) | |
1780 continue; | |
1781 | |
1782 instantiate_virtual_regs_in_insn (insn); | |
1783 | |
1784 if (INSN_DELETED_P (insn)) | |
1785 continue; | |
1786 | |
1787 for_each_rtx (®_NOTES (insn), instantiate_virtual_regs_in_rtx, NULL); | |
1788 | |
1789 /* Instantiate any virtual registers in CALL_INSN_FUNCTION_USAGE. */ | |
1790 if (GET_CODE (insn) == CALL_INSN) | |
1791 for_each_rtx (&CALL_INSN_FUNCTION_USAGE (insn), | |
1792 instantiate_virtual_regs_in_rtx, NULL); | |
1793 } | |
1794 | |
1795 /* Instantiate the virtual registers in the DECLs for debugging purposes. */ | |
1796 instantiate_decls (current_function_decl); | |
1797 | |
1798 targetm.instantiate_decls (); | |
1799 | |
1800 /* Indicate that, from now on, assign_stack_local should use | |
1801 frame_pointer_rtx. */ | |
1802 virtuals_instantiated = 1; | |
1803 return 0; | |
1804 } | |
1805 | |
1806 struct rtl_opt_pass pass_instantiate_virtual_regs = | |
1807 { | |
1808 { | |
1809 RTL_PASS, | |
1810 "vregs", /* name */ | |
1811 NULL, /* gate */ | |
1812 instantiate_virtual_regs, /* execute */ | |
1813 NULL, /* sub */ | |
1814 NULL, /* next */ | |
1815 0, /* static_pass_number */ | |
1816 0, /* tv_id */ | |
1817 0, /* properties_required */ | |
1818 0, /* properties_provided */ | |
1819 0, /* properties_destroyed */ | |
1820 0, /* todo_flags_start */ | |
1821 TODO_dump_func /* todo_flags_finish */ | |
1822 } | |
1823 }; | |
1824 | |
1825 | |
1826 /* Return 1 if EXP is an aggregate type (or a value with aggregate type). | |
1827 This means a type for which function calls must pass an address to the | |
1828 function or get an address back from the function. | |
1829 EXP may be a type node or an expression (whose type is tested). */ | |
1830 | |
1831 int | |
1832 aggregate_value_p (const_tree exp, const_tree fntype) | |
1833 { | |
1834 int i, regno, nregs; | |
1835 rtx reg; | |
1836 | |
1837 const_tree type = (TYPE_P (exp)) ? exp : TREE_TYPE (exp); | |
1838 | |
1839 /* DECL node associated with FNTYPE when relevant, which we might need to | |
1840 check for by-invisible-reference returns, typically for CALL_EXPR input | |
1841 EXPressions. */ | |
1842 const_tree fndecl = NULL_TREE; | |
1843 | |
1844 if (fntype) | |
1845 switch (TREE_CODE (fntype)) | |
1846 { | |
1847 case CALL_EXPR: | |
1848 fndecl = get_callee_fndecl (fntype); | |
1849 fntype = (fndecl | |
1850 ? TREE_TYPE (fndecl) | |
1851 : TREE_TYPE (TREE_TYPE (CALL_EXPR_FN (fntype)))); | |
1852 break; | |
1853 case FUNCTION_DECL: | |
1854 fndecl = fntype; | |
1855 fntype = TREE_TYPE (fndecl); | |
1856 break; | |
1857 case FUNCTION_TYPE: | |
1858 case METHOD_TYPE: | |
1859 break; | |
1860 case IDENTIFIER_NODE: | |
1861 fntype = 0; | |
1862 break; | |
1863 default: | |
1864 /* We don't expect other rtl types here. */ | |
1865 gcc_unreachable (); | |
1866 } | |
1867 | |
1868 if (TREE_CODE (type) == VOID_TYPE) | |
1869 return 0; | |
1870 | |
1871 /* If the front end has decided that this needs to be passed by | |
1872 reference, do so. */ | |
1873 if ((TREE_CODE (exp) == PARM_DECL || TREE_CODE (exp) == RESULT_DECL) | |
1874 && DECL_BY_REFERENCE (exp)) | |
1875 return 1; | |
1876 | |
1877 /* If the EXPression is a CALL_EXPR, honor DECL_BY_REFERENCE set on the | |
1878 called function RESULT_DECL, meaning the function returns in memory by | |
1879 invisible reference. This check lets front-ends not set TREE_ADDRESSABLE | |
1880 on the function type, which used to be the way to request such a return | |
1881 mechanism but might now be causing troubles at gimplification time if | |
1882 temporaries with the function type need to be created. */ | |
1883 if (TREE_CODE (exp) == CALL_EXPR && fndecl && DECL_RESULT (fndecl) | |
1884 && DECL_BY_REFERENCE (DECL_RESULT (fndecl))) | |
1885 return 1; | |
1886 | |
1887 if (targetm.calls.return_in_memory (type, fntype)) | |
1888 return 1; | |
1889 /* Types that are TREE_ADDRESSABLE must be constructed in memory, | |
1890 and thus can't be returned in registers. */ | |
1891 if (TREE_ADDRESSABLE (type)) | |
1892 return 1; | |
1893 if (flag_pcc_struct_return && AGGREGATE_TYPE_P (type)) | |
1894 return 1; | |
1895 /* Make sure we have suitable call-clobbered regs to return | |
1896 the value in; if not, we must return it in memory. */ | |
1897 reg = hard_function_value (type, 0, fntype, 0); | |
1898 | |
1899 /* If we have something other than a REG (e.g. a PARALLEL), then assume | |
1900 it is OK. */ | |
1901 if (!REG_P (reg)) | |
1902 return 0; | |
1903 | |
1904 regno = REGNO (reg); | |
1905 nregs = hard_regno_nregs[regno][TYPE_MODE (type)]; | |
1906 for (i = 0; i < nregs; i++) | |
1907 if (! call_used_regs[regno + i]) | |
1908 return 1; | |
1909 return 0; | |
1910 } | |
1911 | |
1912 /* Return true if we should assign DECL a pseudo register; false if it | |
1913 should live on the local stack. */ | |
1914 | |
1915 bool | |
1916 use_register_for_decl (const_tree decl) | |
1917 { | |
1918 if (!targetm.calls.allocate_stack_slots_for_args()) | |
1919 return true; | |
1920 | |
1921 /* Honor volatile. */ | |
1922 if (TREE_SIDE_EFFECTS (decl)) | |
1923 return false; | |
1924 | |
1925 /* Honor addressability. */ | |
1926 if (TREE_ADDRESSABLE (decl)) | |
1927 return false; | |
1928 | |
1929 /* Only register-like things go in registers. */ | |
1930 if (DECL_MODE (decl) == BLKmode) | |
1931 return false; | |
1932 | |
1933 /* If -ffloat-store specified, don't put explicit float variables | |
1934 into registers. */ | |
1935 /* ??? This should be checked after DECL_ARTIFICIAL, but tree-ssa | |
1936 propagates values across these stores, and it probably shouldn't. */ | |
1937 if (flag_float_store && FLOAT_TYPE_P (TREE_TYPE (decl))) | |
1938 return false; | |
1939 | |
1940 /* If we're not interested in tracking debugging information for | |
1941 this decl, then we can certainly put it in a register. */ | |
1942 if (DECL_IGNORED_P (decl)) | |
1943 return true; | |
1944 | |
1945 if (optimize) | |
1946 return true; | |
1947 | |
1948 if (!DECL_REGISTER (decl)) | |
1949 return false; | |
1950 | |
1951 switch (TREE_CODE (TREE_TYPE (decl))) | |
1952 { | |
1953 case RECORD_TYPE: | |
1954 case UNION_TYPE: | |
1955 case QUAL_UNION_TYPE: | |
1956 /* When not optimizing, disregard register keyword for variables with | |
1957 types containing methods, otherwise the methods won't be callable | |
1958 from the debugger. */ | |
1959 if (TYPE_METHODS (TREE_TYPE (decl))) | |
1960 return false; | |
1961 break; | |
1962 default: | |
1963 break; | |
1964 } | |
1965 | |
1966 return true; | |
1967 } | |
1968 | |
1969 /* Return true if TYPE should be passed by invisible reference. */ | |
1970 | |
1971 bool | |
1972 pass_by_reference (CUMULATIVE_ARGS *ca, enum machine_mode mode, | |
1973 tree type, bool named_arg) | |
1974 { | |
1975 if (type) | |
1976 { | |
1977 /* If this type contains non-trivial constructors, then it is | |
1978 forbidden for the middle-end to create any new copies. */ | |
1979 if (TREE_ADDRESSABLE (type)) | |
1980 return true; | |
1981 | |
1982 /* GCC post 3.4 passes *all* variable sized types by reference. */ | |
1983 if (!TYPE_SIZE (type) || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
1984 return true; | |
1985 } | |
1986 | |
1987 return targetm.calls.pass_by_reference (ca, mode, type, named_arg); | |
1988 } | |
1989 | |
1990 /* Return true if TYPE, which is passed by reference, should be callee | |
1991 copied instead of caller copied. */ | |
1992 | |
1993 bool | |
1994 reference_callee_copied (CUMULATIVE_ARGS *ca, enum machine_mode mode, | |
1995 tree type, bool named_arg) | |
1996 { | |
1997 if (type && TREE_ADDRESSABLE (type)) | |
1998 return false; | |
1999 return targetm.calls.callee_copies (ca, mode, type, named_arg); | |
2000 } | |
2001 | |
2002 /* Structures to communicate between the subroutines of assign_parms. | |
2003 The first holds data persistent across all parameters, the second | |
2004 is cleared out for each parameter. */ | |
2005 | |
2006 struct assign_parm_data_all | |
2007 { | |
2008 CUMULATIVE_ARGS args_so_far; | |
2009 struct args_size stack_args_size; | |
2010 tree function_result_decl; | |
2011 tree orig_fnargs; | |
2012 rtx first_conversion_insn; | |
2013 rtx last_conversion_insn; | |
2014 HOST_WIDE_INT pretend_args_size; | |
2015 HOST_WIDE_INT extra_pretend_bytes; | |
2016 int reg_parm_stack_space; | |
2017 }; | |
2018 | |
2019 struct assign_parm_data_one | |
2020 { | |
2021 tree nominal_type; | |
2022 tree passed_type; | |
2023 rtx entry_parm; | |
2024 rtx stack_parm; | |
2025 enum machine_mode nominal_mode; | |
2026 enum machine_mode passed_mode; | |
2027 enum machine_mode promoted_mode; | |
2028 struct locate_and_pad_arg_data locate; | |
2029 int partial; | |
2030 BOOL_BITFIELD named_arg : 1; | |
2031 BOOL_BITFIELD passed_pointer : 1; | |
2032 BOOL_BITFIELD on_stack : 1; | |
2033 BOOL_BITFIELD loaded_in_reg : 1; | |
2034 }; | |
2035 | |
2036 /* A subroutine of assign_parms. Initialize ALL. */ | |
2037 | |
2038 static void | |
2039 assign_parms_initialize_all (struct assign_parm_data_all *all) | |
2040 { | |
2041 tree fntype; | |
2042 | |
2043 memset (all, 0, sizeof (*all)); | |
2044 | |
2045 fntype = TREE_TYPE (current_function_decl); | |
2046 | |
2047 #ifdef INIT_CUMULATIVE_INCOMING_ARGS | |
2048 INIT_CUMULATIVE_INCOMING_ARGS (all->args_so_far, fntype, NULL_RTX); | |
2049 #else | |
2050 INIT_CUMULATIVE_ARGS (all->args_so_far, fntype, NULL_RTX, | |
2051 current_function_decl, -1); | |
2052 #endif | |
2053 | |
2054 #ifdef REG_PARM_STACK_SPACE | |
2055 all->reg_parm_stack_space = REG_PARM_STACK_SPACE (current_function_decl); | |
2056 #endif | |
2057 } | |
2058 | |
2059 /* If ARGS contains entries with complex types, split the entry into two | |
2060 entries of the component type. Return a new list of substitutions are | |
2061 needed, else the old list. */ | |
2062 | |
2063 static tree | |
2064 split_complex_args (tree args) | |
2065 { | |
2066 tree p; | |
2067 | |
2068 /* Before allocating memory, check for the common case of no complex. */ | |
2069 for (p = args; p; p = TREE_CHAIN (p)) | |
2070 { | |
2071 tree type = TREE_TYPE (p); | |
2072 if (TREE_CODE (type) == COMPLEX_TYPE | |
2073 && targetm.calls.split_complex_arg (type)) | |
2074 goto found; | |
2075 } | |
2076 return args; | |
2077 | |
2078 found: | |
2079 args = copy_list (args); | |
2080 | |
2081 for (p = args; p; p = TREE_CHAIN (p)) | |
2082 { | |
2083 tree type = TREE_TYPE (p); | |
2084 if (TREE_CODE (type) == COMPLEX_TYPE | |
2085 && targetm.calls.split_complex_arg (type)) | |
2086 { | |
2087 tree decl; | |
2088 tree subtype = TREE_TYPE (type); | |
2089 bool addressable = TREE_ADDRESSABLE (p); | |
2090 | |
2091 /* Rewrite the PARM_DECL's type with its component. */ | |
2092 TREE_TYPE (p) = subtype; | |
2093 DECL_ARG_TYPE (p) = TREE_TYPE (DECL_ARG_TYPE (p)); | |
2094 DECL_MODE (p) = VOIDmode; | |
2095 DECL_SIZE (p) = NULL; | |
2096 DECL_SIZE_UNIT (p) = NULL; | |
2097 /* If this arg must go in memory, put it in a pseudo here. | |
2098 We can't allow it to go in memory as per normal parms, | |
2099 because the usual place might not have the imag part | |
2100 adjacent to the real part. */ | |
2101 DECL_ARTIFICIAL (p) = addressable; | |
2102 DECL_IGNORED_P (p) = addressable; | |
2103 TREE_ADDRESSABLE (p) = 0; | |
2104 layout_decl (p, 0); | |
2105 | |
2106 /* Build a second synthetic decl. */ | |
2107 decl = build_decl (PARM_DECL, NULL_TREE, subtype); | |
2108 DECL_ARG_TYPE (decl) = DECL_ARG_TYPE (p); | |
2109 DECL_ARTIFICIAL (decl) = addressable; | |
2110 DECL_IGNORED_P (decl) = addressable; | |
2111 layout_decl (decl, 0); | |
2112 | |
2113 /* Splice it in; skip the new decl. */ | |
2114 TREE_CHAIN (decl) = TREE_CHAIN (p); | |
2115 TREE_CHAIN (p) = decl; | |
2116 p = decl; | |
2117 } | |
2118 } | |
2119 | |
2120 return args; | |
2121 } | |
2122 | |
2123 /* A subroutine of assign_parms. Adjust the parameter list to incorporate | |
2124 the hidden struct return argument, and (abi willing) complex args. | |
2125 Return the new parameter list. */ | |
2126 | |
2127 static tree | |
2128 assign_parms_augmented_arg_list (struct assign_parm_data_all *all) | |
2129 { | |
2130 tree fndecl = current_function_decl; | |
2131 tree fntype = TREE_TYPE (fndecl); | |
2132 tree fnargs = DECL_ARGUMENTS (fndecl); | |
2133 | |
2134 /* If struct value address is treated as the first argument, make it so. */ | |
2135 if (aggregate_value_p (DECL_RESULT (fndecl), fndecl) | |
2136 && ! cfun->returns_pcc_struct | |
2137 && targetm.calls.struct_value_rtx (TREE_TYPE (fndecl), 1) == 0) | |
2138 { | |
2139 tree type = build_pointer_type (TREE_TYPE (fntype)); | |
2140 tree decl; | |
2141 | |
2142 decl = build_decl (PARM_DECL, NULL_TREE, type); | |
2143 DECL_ARG_TYPE (decl) = type; | |
2144 DECL_ARTIFICIAL (decl) = 1; | |
2145 DECL_IGNORED_P (decl) = 1; | |
2146 | |
2147 TREE_CHAIN (decl) = fnargs; | |
2148 fnargs = decl; | |
2149 all->function_result_decl = decl; | |
2150 } | |
2151 | |
2152 all->orig_fnargs = fnargs; | |
2153 | |
2154 /* If the target wants to split complex arguments into scalars, do so. */ | |
2155 if (targetm.calls.split_complex_arg) | |
2156 fnargs = split_complex_args (fnargs); | |
2157 | |
2158 return fnargs; | |
2159 } | |
2160 | |
2161 /* A subroutine of assign_parms. Examine PARM and pull out type and mode | |
2162 data for the parameter. Incorporate ABI specifics such as pass-by- | |
2163 reference and type promotion. */ | |
2164 | |
2165 static void | |
2166 assign_parm_find_data_types (struct assign_parm_data_all *all, tree parm, | |
2167 struct assign_parm_data_one *data) | |
2168 { | |
2169 tree nominal_type, passed_type; | |
2170 enum machine_mode nominal_mode, passed_mode, promoted_mode; | |
2171 | |
2172 memset (data, 0, sizeof (*data)); | |
2173 | |
2174 /* NAMED_ARG is a misnomer. We really mean 'non-variadic'. */ | |
2175 if (!cfun->stdarg) | |
2176 data->named_arg = 1; /* No variadic parms. */ | |
2177 else if (TREE_CHAIN (parm)) | |
2178 data->named_arg = 1; /* Not the last non-variadic parm. */ | |
2179 else if (targetm.calls.strict_argument_naming (&all->args_so_far)) | |
2180 data->named_arg = 1; /* Only variadic ones are unnamed. */ | |
2181 else | |
2182 data->named_arg = 0; /* Treat as variadic. */ | |
2183 | |
2184 nominal_type = TREE_TYPE (parm); | |
2185 passed_type = DECL_ARG_TYPE (parm); | |
2186 | |
2187 /* Look out for errors propagating this far. Also, if the parameter's | |
2188 type is void then its value doesn't matter. */ | |
2189 if (TREE_TYPE (parm) == error_mark_node | |
2190 /* This can happen after weird syntax errors | |
2191 or if an enum type is defined among the parms. */ | |
2192 || TREE_CODE (parm) != PARM_DECL | |
2193 || passed_type == NULL | |
2194 || VOID_TYPE_P (nominal_type)) | |
2195 { | |
2196 nominal_type = passed_type = void_type_node; | |
2197 nominal_mode = passed_mode = promoted_mode = VOIDmode; | |
2198 goto egress; | |
2199 } | |
2200 | |
2201 /* Find mode of arg as it is passed, and mode of arg as it should be | |
2202 during execution of this function. */ | |
2203 passed_mode = TYPE_MODE (passed_type); | |
2204 nominal_mode = TYPE_MODE (nominal_type); | |
2205 | |
2206 /* If the parm is to be passed as a transparent union, use the type of | |
2207 the first field for the tests below. We have already verified that | |
2208 the modes are the same. */ | |
2209 if (TREE_CODE (passed_type) == UNION_TYPE | |
2210 && TYPE_TRANSPARENT_UNION (passed_type)) | |
2211 passed_type = TREE_TYPE (TYPE_FIELDS (passed_type)); | |
2212 | |
2213 /* See if this arg was passed by invisible reference. */ | |
2214 if (pass_by_reference (&all->args_so_far, passed_mode, | |
2215 passed_type, data->named_arg)) | |
2216 { | |
2217 passed_type = nominal_type = build_pointer_type (passed_type); | |
2218 data->passed_pointer = true; | |
2219 passed_mode = nominal_mode = Pmode; | |
2220 } | |
2221 | |
2222 /* Find mode as it is passed by the ABI. */ | |
2223 promoted_mode = passed_mode; | |
2224 if (targetm.calls.promote_function_args (TREE_TYPE (current_function_decl))) | |
2225 { | |
2226 int unsignedp = TYPE_UNSIGNED (passed_type); | |
2227 promoted_mode = promote_mode (passed_type, promoted_mode, | |
2228 &unsignedp, 1); | |
2229 } | |
2230 | |
2231 egress: | |
2232 data->nominal_type = nominal_type; | |
2233 data->passed_type = passed_type; | |
2234 data->nominal_mode = nominal_mode; | |
2235 data->passed_mode = passed_mode; | |
2236 data->promoted_mode = promoted_mode; | |
2237 } | |
2238 | |
2239 /* A subroutine of assign_parms. Invoke setup_incoming_varargs. */ | |
2240 | |
2241 static void | |
2242 assign_parms_setup_varargs (struct assign_parm_data_all *all, | |
2243 struct assign_parm_data_one *data, bool no_rtl) | |
2244 { | |
2245 int varargs_pretend_bytes = 0; | |
2246 | |
2247 targetm.calls.setup_incoming_varargs (&all->args_so_far, | |
2248 data->promoted_mode, | |
2249 data->passed_type, | |
2250 &varargs_pretend_bytes, no_rtl); | |
2251 | |
2252 /* If the back-end has requested extra stack space, record how much is | |
2253 needed. Do not change pretend_args_size otherwise since it may be | |
2254 nonzero from an earlier partial argument. */ | |
2255 if (varargs_pretend_bytes > 0) | |
2256 all->pretend_args_size = varargs_pretend_bytes; | |
2257 } | |
2258 | |
2259 /* A subroutine of assign_parms. Set DATA->ENTRY_PARM corresponding to | |
2260 the incoming location of the current parameter. */ | |
2261 | |
2262 static void | |
2263 assign_parm_find_entry_rtl (struct assign_parm_data_all *all, | |
2264 struct assign_parm_data_one *data) | |
2265 { | |
2266 HOST_WIDE_INT pretend_bytes = 0; | |
2267 rtx entry_parm; | |
2268 bool in_regs; | |
2269 | |
2270 if (data->promoted_mode == VOIDmode) | |
2271 { | |
2272 data->entry_parm = data->stack_parm = const0_rtx; | |
2273 return; | |
2274 } | |
2275 | |
2276 #ifdef FUNCTION_INCOMING_ARG | |
2277 entry_parm = FUNCTION_INCOMING_ARG (all->args_so_far, data->promoted_mode, | |
2278 data->passed_type, data->named_arg); | |
2279 #else | |
2280 entry_parm = FUNCTION_ARG (all->args_so_far, data->promoted_mode, | |
2281 data->passed_type, data->named_arg); | |
2282 #endif | |
2283 | |
2284 if (entry_parm == 0) | |
2285 data->promoted_mode = data->passed_mode; | |
2286 | |
2287 /* Determine parm's home in the stack, in case it arrives in the stack | |
2288 or we should pretend it did. Compute the stack position and rtx where | |
2289 the argument arrives and its size. | |
2290 | |
2291 There is one complexity here: If this was a parameter that would | |
2292 have been passed in registers, but wasn't only because it is | |
2293 __builtin_va_alist, we want locate_and_pad_parm to treat it as if | |
2294 it came in a register so that REG_PARM_STACK_SPACE isn't skipped. | |
2295 In this case, we call FUNCTION_ARG with NAMED set to 1 instead of 0 | |
2296 as it was the previous time. */ | |
2297 in_regs = entry_parm != 0; | |
2298 #ifdef STACK_PARMS_IN_REG_PARM_AREA | |
2299 in_regs = true; | |
2300 #endif | |
2301 if (!in_regs && !data->named_arg) | |
2302 { | |
2303 if (targetm.calls.pretend_outgoing_varargs_named (&all->args_so_far)) | |
2304 { | |
2305 rtx tem; | |
2306 #ifdef FUNCTION_INCOMING_ARG | |
2307 tem = FUNCTION_INCOMING_ARG (all->args_so_far, data->promoted_mode, | |
2308 data->passed_type, true); | |
2309 #else | |
2310 tem = FUNCTION_ARG (all->args_so_far, data->promoted_mode, | |
2311 data->passed_type, true); | |
2312 #endif | |
2313 in_regs = tem != NULL; | |
2314 } | |
2315 } | |
2316 | |
2317 /* If this parameter was passed both in registers and in the stack, use | |
2318 the copy on the stack. */ | |
2319 if (targetm.calls.must_pass_in_stack (data->promoted_mode, | |
2320 data->passed_type)) | |
2321 entry_parm = 0; | |
2322 | |
2323 if (entry_parm) | |
2324 { | |
2325 int partial; | |
2326 | |
2327 partial = targetm.calls.arg_partial_bytes (&all->args_so_far, | |
2328 data->promoted_mode, | |
2329 data->passed_type, | |
2330 data->named_arg); | |
2331 data->partial = partial; | |
2332 | |
2333 /* The caller might already have allocated stack space for the | |
2334 register parameters. */ | |
2335 if (partial != 0 && all->reg_parm_stack_space == 0) | |
2336 { | |
2337 /* Part of this argument is passed in registers and part | |
2338 is passed on the stack. Ask the prologue code to extend | |
2339 the stack part so that we can recreate the full value. | |
2340 | |
2341 PRETEND_BYTES is the size of the registers we need to store. | |
2342 CURRENT_FUNCTION_PRETEND_ARGS_SIZE is the amount of extra | |
2343 stack space that the prologue should allocate. | |
2344 | |
2345 Internally, gcc assumes that the argument pointer is aligned | |
2346 to STACK_BOUNDARY bits. This is used both for alignment | |
2347 optimizations (see init_emit) and to locate arguments that are | |
2348 aligned to more than PARM_BOUNDARY bits. We must preserve this | |
2349 invariant by rounding CURRENT_FUNCTION_PRETEND_ARGS_SIZE up to | |
2350 a stack boundary. */ | |
2351 | |
2352 /* We assume at most one partial arg, and it must be the first | |
2353 argument on the stack. */ | |
2354 gcc_assert (!all->extra_pretend_bytes && !all->pretend_args_size); | |
2355 | |
2356 pretend_bytes = partial; | |
2357 all->pretend_args_size = CEIL_ROUND (pretend_bytes, STACK_BYTES); | |
2358 | |
2359 /* We want to align relative to the actual stack pointer, so | |
2360 don't include this in the stack size until later. */ | |
2361 all->extra_pretend_bytes = all->pretend_args_size; | |
2362 } | |
2363 } | |
2364 | |
2365 locate_and_pad_parm (data->promoted_mode, data->passed_type, in_regs, | |
2366 entry_parm ? data->partial : 0, current_function_decl, | |
2367 &all->stack_args_size, &data->locate); | |
2368 | |
2369 /* Update parm_stack_boundary if this parameter is passed in the | |
2370 stack. */ | |
2371 if (!in_regs && crtl->parm_stack_boundary < data->locate.boundary) | |
2372 crtl->parm_stack_boundary = data->locate.boundary; | |
2373 | |
2374 /* Adjust offsets to include the pretend args. */ | |
2375 pretend_bytes = all->extra_pretend_bytes - pretend_bytes; | |
2376 data->locate.slot_offset.constant += pretend_bytes; | |
2377 data->locate.offset.constant += pretend_bytes; | |
2378 | |
2379 data->entry_parm = entry_parm; | |
2380 } | |
2381 | |
2382 /* A subroutine of assign_parms. If there is actually space on the stack | |
2383 for this parm, count it in stack_args_size and return true. */ | |
2384 | |
2385 static bool | |
2386 assign_parm_is_stack_parm (struct assign_parm_data_all *all, | |
2387 struct assign_parm_data_one *data) | |
2388 { | |
2389 /* Trivially true if we've no incoming register. */ | |
2390 if (data->entry_parm == NULL) | |
2391 ; | |
2392 /* Also true if we're partially in registers and partially not, | |
2393 since we've arranged to drop the entire argument on the stack. */ | |
2394 else if (data->partial != 0) | |
2395 ; | |
2396 /* Also true if the target says that it's passed in both registers | |
2397 and on the stack. */ | |
2398 else if (GET_CODE (data->entry_parm) == PARALLEL | |
2399 && XEXP (XVECEXP (data->entry_parm, 0, 0), 0) == NULL_RTX) | |
2400 ; | |
2401 /* Also true if the target says that there's stack allocated for | |
2402 all register parameters. */ | |
2403 else if (all->reg_parm_stack_space > 0) | |
2404 ; | |
2405 /* Otherwise, no, this parameter has no ABI defined stack slot. */ | |
2406 else | |
2407 return false; | |
2408 | |
2409 all->stack_args_size.constant += data->locate.size.constant; | |
2410 if (data->locate.size.var) | |
2411 ADD_PARM_SIZE (all->stack_args_size, data->locate.size.var); | |
2412 | |
2413 return true; | |
2414 } | |
2415 | |
2416 /* A subroutine of assign_parms. Given that this parameter is allocated | |
2417 stack space by the ABI, find it. */ | |
2418 | |
2419 static void | |
2420 assign_parm_find_stack_rtl (tree parm, struct assign_parm_data_one *data) | |
2421 { | |
2422 rtx offset_rtx, stack_parm; | |
2423 unsigned int align, boundary; | |
2424 | |
2425 /* If we're passing this arg using a reg, make its stack home the | |
2426 aligned stack slot. */ | |
2427 if (data->entry_parm) | |
2428 offset_rtx = ARGS_SIZE_RTX (data->locate.slot_offset); | |
2429 else | |
2430 offset_rtx = ARGS_SIZE_RTX (data->locate.offset); | |
2431 | |
2432 stack_parm = crtl->args.internal_arg_pointer; | |
2433 if (offset_rtx != const0_rtx) | |
2434 stack_parm = gen_rtx_PLUS (Pmode, stack_parm, offset_rtx); | |
2435 stack_parm = gen_rtx_MEM (data->promoted_mode, stack_parm); | |
2436 | |
2437 set_mem_attributes (stack_parm, parm, 1); | |
2438 /* set_mem_attributes could set MEM_SIZE to the passed mode's size, | |
2439 while promoted mode's size is needed. */ | |
2440 if (data->promoted_mode != BLKmode | |
2441 && data->promoted_mode != DECL_MODE (parm)) | |
2442 { | |
2443 set_mem_size (stack_parm, GEN_INT (GET_MODE_SIZE (data->promoted_mode))); | |
2444 if (MEM_EXPR (stack_parm) && MEM_OFFSET (stack_parm)) | |
2445 { | |
2446 int offset = subreg_lowpart_offset (DECL_MODE (parm), | |
2447 data->promoted_mode); | |
2448 if (offset) | |
2449 set_mem_offset (stack_parm, | |
2450 plus_constant (MEM_OFFSET (stack_parm), -offset)); | |
2451 } | |
2452 } | |
2453 | |
2454 boundary = data->locate.boundary; | |
2455 align = BITS_PER_UNIT; | |
2456 | |
2457 /* If we're padding upward, we know that the alignment of the slot | |
2458 is FUNCTION_ARG_BOUNDARY. If we're using slot_offset, we're | |
2459 intentionally forcing upward padding. Otherwise we have to come | |
2460 up with a guess at the alignment based on OFFSET_RTX. */ | |
2461 if (data->locate.where_pad != downward || data->entry_parm) | |
2462 align = boundary; | |
2463 else if (GET_CODE (offset_rtx) == CONST_INT) | |
2464 { | |
2465 align = INTVAL (offset_rtx) * BITS_PER_UNIT | boundary; | |
2466 align = align & -align; | |
2467 } | |
2468 set_mem_align (stack_parm, align); | |
2469 | |
2470 if (data->entry_parm) | |
2471 set_reg_attrs_for_parm (data->entry_parm, stack_parm); | |
2472 | |
2473 data->stack_parm = stack_parm; | |
2474 } | |
2475 | |
2476 /* A subroutine of assign_parms. Adjust DATA->ENTRY_RTL such that it's | |
2477 always valid and contiguous. */ | |
2478 | |
2479 static void | |
2480 assign_parm_adjust_entry_rtl (struct assign_parm_data_one *data) | |
2481 { | |
2482 rtx entry_parm = data->entry_parm; | |
2483 rtx stack_parm = data->stack_parm; | |
2484 | |
2485 /* If this parm was passed part in regs and part in memory, pretend it | |
2486 arrived entirely in memory by pushing the register-part onto the stack. | |
2487 In the special case of a DImode or DFmode that is split, we could put | |
2488 it together in a pseudoreg directly, but for now that's not worth | |
2489 bothering with. */ | |
2490 if (data->partial != 0) | |
2491 { | |
2492 /* Handle calls that pass values in multiple non-contiguous | |
2493 locations. The Irix 6 ABI has examples of this. */ | |
2494 if (GET_CODE (entry_parm) == PARALLEL) | |
2495 emit_group_store (validize_mem (stack_parm), entry_parm, | |
2496 data->passed_type, | |
2497 int_size_in_bytes (data->passed_type)); | |
2498 else | |
2499 { | |
2500 gcc_assert (data->partial % UNITS_PER_WORD == 0); | |
2501 move_block_from_reg (REGNO (entry_parm), validize_mem (stack_parm), | |
2502 data->partial / UNITS_PER_WORD); | |
2503 } | |
2504 | |
2505 entry_parm = stack_parm; | |
2506 } | |
2507 | |
2508 /* If we didn't decide this parm came in a register, by default it came | |
2509 on the stack. */ | |
2510 else if (entry_parm == NULL) | |
2511 entry_parm = stack_parm; | |
2512 | |
2513 /* When an argument is passed in multiple locations, we can't make use | |
2514 of this information, but we can save some copying if the whole argument | |
2515 is passed in a single register. */ | |
2516 else if (GET_CODE (entry_parm) == PARALLEL | |
2517 && data->nominal_mode != BLKmode | |
2518 && data->passed_mode != BLKmode) | |
2519 { | |
2520 size_t i, len = XVECLEN (entry_parm, 0); | |
2521 | |
2522 for (i = 0; i < len; i++) | |
2523 if (XEXP (XVECEXP (entry_parm, 0, i), 0) != NULL_RTX | |
2524 && REG_P (XEXP (XVECEXP (entry_parm, 0, i), 0)) | |
2525 && (GET_MODE (XEXP (XVECEXP (entry_parm, 0, i), 0)) | |
2526 == data->passed_mode) | |
2527 && INTVAL (XEXP (XVECEXP (entry_parm, 0, i), 1)) == 0) | |
2528 { | |
2529 entry_parm = XEXP (XVECEXP (entry_parm, 0, i), 0); | |
2530 break; | |
2531 } | |
2532 } | |
2533 | |
2534 data->entry_parm = entry_parm; | |
2535 } | |
2536 | |
2537 /* A subroutine of assign_parms. Reconstitute any values which were | |
2538 passed in multiple registers and would fit in a single register. */ | |
2539 | |
2540 static void | |
2541 assign_parm_remove_parallels (struct assign_parm_data_one *data) | |
2542 { | |
2543 rtx entry_parm = data->entry_parm; | |
2544 | |
2545 /* Convert the PARALLEL to a REG of the same mode as the parallel. | |
2546 This can be done with register operations rather than on the | |
2547 stack, even if we will store the reconstituted parameter on the | |
2548 stack later. */ | |
2549 if (GET_CODE (entry_parm) == PARALLEL && GET_MODE (entry_parm) != BLKmode) | |
2550 { | |
2551 rtx parmreg = gen_reg_rtx (GET_MODE (entry_parm)); | |
2552 emit_group_store (parmreg, entry_parm, data->passed_type, | |
2553 GET_MODE_SIZE (GET_MODE (entry_parm))); | |
2554 entry_parm = parmreg; | |
2555 } | |
2556 | |
2557 data->entry_parm = entry_parm; | |
2558 } | |
2559 | |
2560 /* A subroutine of assign_parms. Adjust DATA->STACK_RTL such that it's | |
2561 always valid and properly aligned. */ | |
2562 | |
2563 static void | |
2564 assign_parm_adjust_stack_rtl (struct assign_parm_data_one *data) | |
2565 { | |
2566 rtx stack_parm = data->stack_parm; | |
2567 | |
2568 /* If we can't trust the parm stack slot to be aligned enough for its | |
2569 ultimate type, don't use that slot after entry. We'll make another | |
2570 stack slot, if we need one. */ | |
2571 if (stack_parm | |
2572 && ((STRICT_ALIGNMENT | |
2573 && GET_MODE_ALIGNMENT (data->nominal_mode) > MEM_ALIGN (stack_parm)) | |
2574 || (data->nominal_type | |
2575 && TYPE_ALIGN (data->nominal_type) > MEM_ALIGN (stack_parm) | |
2576 && MEM_ALIGN (stack_parm) < PREFERRED_STACK_BOUNDARY))) | |
2577 stack_parm = NULL; | |
2578 | |
2579 /* If parm was passed in memory, and we need to convert it on entry, | |
2580 don't store it back in that same slot. */ | |
2581 else if (data->entry_parm == stack_parm | |
2582 && data->nominal_mode != BLKmode | |
2583 && data->nominal_mode != data->passed_mode) | |
2584 stack_parm = NULL; | |
2585 | |
2586 /* If stack protection is in effect for this function, don't leave any | |
2587 pointers in their passed stack slots. */ | |
2588 else if (crtl->stack_protect_guard | |
2589 && (flag_stack_protect == 2 | |
2590 || data->passed_pointer | |
2591 || POINTER_TYPE_P (data->nominal_type))) | |
2592 stack_parm = NULL; | |
2593 | |
2594 data->stack_parm = stack_parm; | |
2595 } | |
2596 | |
2597 /* A subroutine of assign_parms. Return true if the current parameter | |
2598 should be stored as a BLKmode in the current frame. */ | |
2599 | |
2600 static bool | |
2601 assign_parm_setup_block_p (struct assign_parm_data_one *data) | |
2602 { | |
2603 if (data->nominal_mode == BLKmode) | |
2604 return true; | |
2605 if (GET_MODE (data->entry_parm) == BLKmode) | |
2606 return true; | |
2607 | |
2608 #ifdef BLOCK_REG_PADDING | |
2609 /* Only assign_parm_setup_block knows how to deal with register arguments | |
2610 that are padded at the least significant end. */ | |
2611 if (REG_P (data->entry_parm) | |
2612 && GET_MODE_SIZE (data->promoted_mode) < UNITS_PER_WORD | |
2613 && (BLOCK_REG_PADDING (data->passed_mode, data->passed_type, 1) | |
2614 == (BYTES_BIG_ENDIAN ? upward : downward))) | |
2615 return true; | |
2616 #endif | |
2617 | |
2618 return false; | |
2619 } | |
2620 | |
2621 /* A subroutine of assign_parms. Arrange for the parameter to be | |
2622 present and valid in DATA->STACK_RTL. */ | |
2623 | |
2624 static void | |
2625 assign_parm_setup_block (struct assign_parm_data_all *all, | |
2626 tree parm, struct assign_parm_data_one *data) | |
2627 { | |
2628 rtx entry_parm = data->entry_parm; | |
2629 rtx stack_parm = data->stack_parm; | |
2630 HOST_WIDE_INT size; | |
2631 HOST_WIDE_INT size_stored; | |
2632 | |
2633 if (GET_CODE (entry_parm) == PARALLEL) | |
2634 entry_parm = emit_group_move_into_temps (entry_parm); | |
2635 | |
2636 size = int_size_in_bytes (data->passed_type); | |
2637 size_stored = CEIL_ROUND (size, UNITS_PER_WORD); | |
2638 if (stack_parm == 0) | |
2639 { | |
2640 DECL_ALIGN (parm) = MAX (DECL_ALIGN (parm), BITS_PER_WORD); | |
2641 stack_parm = assign_stack_local (BLKmode, size_stored, | |
2642 DECL_ALIGN (parm)); | |
2643 if (GET_MODE_SIZE (GET_MODE (entry_parm)) == size) | |
2644 PUT_MODE (stack_parm, GET_MODE (entry_parm)); | |
2645 set_mem_attributes (stack_parm, parm, 1); | |
2646 } | |
2647 | |
2648 /* If a BLKmode arrives in registers, copy it to a stack slot. Handle | |
2649 calls that pass values in multiple non-contiguous locations. */ | |
2650 if (REG_P (entry_parm) || GET_CODE (entry_parm) == PARALLEL) | |
2651 { | |
2652 rtx mem; | |
2653 | |
2654 /* Note that we will be storing an integral number of words. | |
2655 So we have to be careful to ensure that we allocate an | |
2656 integral number of words. We do this above when we call | |
2657 assign_stack_local if space was not allocated in the argument | |
2658 list. If it was, this will not work if PARM_BOUNDARY is not | |
2659 a multiple of BITS_PER_WORD. It isn't clear how to fix this | |
2660 if it becomes a problem. Exception is when BLKmode arrives | |
2661 with arguments not conforming to word_mode. */ | |
2662 | |
2663 if (data->stack_parm == 0) | |
2664 ; | |
2665 else if (GET_CODE (entry_parm) == PARALLEL) | |
2666 ; | |
2667 else | |
2668 gcc_assert (!size || !(PARM_BOUNDARY % BITS_PER_WORD)); | |
2669 | |
2670 mem = validize_mem (stack_parm); | |
2671 | |
2672 /* Handle values in multiple non-contiguous locations. */ | |
2673 if (GET_CODE (entry_parm) == PARALLEL) | |
2674 { | |
2675 push_to_sequence2 (all->first_conversion_insn, | |
2676 all->last_conversion_insn); | |
2677 emit_group_store (mem, entry_parm, data->passed_type, size); | |
2678 all->first_conversion_insn = get_insns (); | |
2679 all->last_conversion_insn = get_last_insn (); | |
2680 end_sequence (); | |
2681 } | |
2682 | |
2683 else if (size == 0) | |
2684 ; | |
2685 | |
2686 /* If SIZE is that of a mode no bigger than a word, just use | |
2687 that mode's store operation. */ | |
2688 else if (size <= UNITS_PER_WORD) | |
2689 { | |
2690 enum machine_mode mode | |
2691 = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0); | |
2692 | |
2693 if (mode != BLKmode | |
2694 #ifdef BLOCK_REG_PADDING | |
2695 && (size == UNITS_PER_WORD | |
2696 || (BLOCK_REG_PADDING (mode, data->passed_type, 1) | |
2697 != (BYTES_BIG_ENDIAN ? upward : downward))) | |
2698 #endif | |
2699 ) | |
2700 { | |
2701 rtx reg; | |
2702 | |
2703 /* We are really truncating a word_mode value containing | |
2704 SIZE bytes into a value of mode MODE. If such an | |
2705 operation requires no actual instructions, we can refer | |
2706 to the value directly in mode MODE, otherwise we must | |
2707 start with the register in word_mode and explicitly | |
2708 convert it. */ | |
2709 if (TRULY_NOOP_TRUNCATION (size * BITS_PER_UNIT, BITS_PER_WORD)) | |
2710 reg = gen_rtx_REG (mode, REGNO (entry_parm)); | |
2711 else | |
2712 { | |
2713 reg = gen_rtx_REG (word_mode, REGNO (entry_parm)); | |
2714 reg = convert_to_mode (mode, copy_to_reg (reg), 1); | |
2715 } | |
2716 emit_move_insn (change_address (mem, mode, 0), reg); | |
2717 } | |
2718 | |
2719 /* Blocks smaller than a word on a BYTES_BIG_ENDIAN | |
2720 machine must be aligned to the left before storing | |
2721 to memory. Note that the previous test doesn't | |
2722 handle all cases (e.g. SIZE == 3). */ | |
2723 else if (size != UNITS_PER_WORD | |
2724 #ifdef BLOCK_REG_PADDING | |
2725 && (BLOCK_REG_PADDING (mode, data->passed_type, 1) | |
2726 == downward) | |
2727 #else | |
2728 && BYTES_BIG_ENDIAN | |
2729 #endif | |
2730 ) | |
2731 { | |
2732 rtx tem, x; | |
2733 int by = (UNITS_PER_WORD - size) * BITS_PER_UNIT; | |
2734 rtx reg = gen_rtx_REG (word_mode, REGNO (entry_parm)); | |
2735 | |
2736 x = expand_shift (LSHIFT_EXPR, word_mode, reg, | |
2737 build_int_cst (NULL_TREE, by), | |
2738 NULL_RTX, 1); | |
2739 tem = change_address (mem, word_mode, 0); | |
2740 emit_move_insn (tem, x); | |
2741 } | |
2742 else | |
2743 move_block_from_reg (REGNO (entry_parm), mem, | |
2744 size_stored / UNITS_PER_WORD); | |
2745 } | |
2746 else | |
2747 move_block_from_reg (REGNO (entry_parm), mem, | |
2748 size_stored / UNITS_PER_WORD); | |
2749 } | |
2750 else if (data->stack_parm == 0) | |
2751 { | |
2752 push_to_sequence2 (all->first_conversion_insn, all->last_conversion_insn); | |
2753 emit_block_move (stack_parm, data->entry_parm, GEN_INT (size), | |
2754 BLOCK_OP_NORMAL); | |
2755 all->first_conversion_insn = get_insns (); | |
2756 all->last_conversion_insn = get_last_insn (); | |
2757 end_sequence (); | |
2758 } | |
2759 | |
2760 data->stack_parm = stack_parm; | |
2761 SET_DECL_RTL (parm, stack_parm); | |
2762 } | |
2763 | |
2764 /* A subroutine of assign_parms. Allocate a pseudo to hold the current | |
2765 parameter. Get it there. Perform all ABI specified conversions. */ | |
2766 | |
2767 static void | |
2768 assign_parm_setup_reg (struct assign_parm_data_all *all, tree parm, | |
2769 struct assign_parm_data_one *data) | |
2770 { | |
2771 rtx parmreg; | |
2772 enum machine_mode promoted_nominal_mode; | |
2773 int unsignedp = TYPE_UNSIGNED (TREE_TYPE (parm)); | |
2774 bool did_conversion = false; | |
2775 | |
2776 /* Store the parm in a pseudoregister during the function, but we may | |
2777 need to do it in a wider mode. */ | |
2778 | |
2779 /* This is not really promoting for a call. However we need to be | |
2780 consistent with assign_parm_find_data_types and expand_expr_real_1. */ | |
2781 promoted_nominal_mode | |
2782 = promote_mode (data->nominal_type, data->nominal_mode, &unsignedp, 1); | |
2783 | |
2784 parmreg = gen_reg_rtx (promoted_nominal_mode); | |
2785 | |
2786 if (!DECL_ARTIFICIAL (parm)) | |
2787 mark_user_reg (parmreg); | |
2788 | |
2789 /* If this was an item that we received a pointer to, | |
2790 set DECL_RTL appropriately. */ | |
2791 if (data->passed_pointer) | |
2792 { | |
2793 rtx x = gen_rtx_MEM (TYPE_MODE (TREE_TYPE (data->passed_type)), parmreg); | |
2794 set_mem_attributes (x, parm, 1); | |
2795 SET_DECL_RTL (parm, x); | |
2796 } | |
2797 else | |
2798 SET_DECL_RTL (parm, parmreg); | |
2799 | |
2800 assign_parm_remove_parallels (data); | |
2801 | |
2802 /* Copy the value into the register. */ | |
2803 if (data->nominal_mode != data->passed_mode | |
2804 || promoted_nominal_mode != data->promoted_mode) | |
2805 { | |
2806 int save_tree_used; | |
2807 | |
2808 /* ENTRY_PARM has been converted to PROMOTED_MODE, its | |
2809 mode, by the caller. We now have to convert it to | |
2810 NOMINAL_MODE, if different. However, PARMREG may be in | |
2811 a different mode than NOMINAL_MODE if it is being stored | |
2812 promoted. | |
2813 | |
2814 If ENTRY_PARM is a hard register, it might be in a register | |
2815 not valid for operating in its mode (e.g., an odd-numbered | |
2816 register for a DFmode). In that case, moves are the only | |
2817 thing valid, so we can't do a convert from there. This | |
2818 occurs when the calling sequence allow such misaligned | |
2819 usages. | |
2820 | |
2821 In addition, the conversion may involve a call, which could | |
2822 clobber parameters which haven't been copied to pseudo | |
2823 registers yet. Therefore, we must first copy the parm to | |
2824 a pseudo reg here, and save the conversion until after all | |
2825 parameters have been moved. */ | |
2826 | |
2827 rtx tempreg = gen_reg_rtx (GET_MODE (data->entry_parm)); | |
2828 | |
2829 emit_move_insn (tempreg, validize_mem (data->entry_parm)); | |
2830 | |
2831 push_to_sequence2 (all->first_conversion_insn, all->last_conversion_insn); | |
2832 tempreg = convert_to_mode (data->nominal_mode, tempreg, unsignedp); | |
2833 | |
2834 if (GET_CODE (tempreg) == SUBREG | |
2835 && GET_MODE (tempreg) == data->nominal_mode | |
2836 && REG_P (SUBREG_REG (tempreg)) | |
2837 && data->nominal_mode == data->passed_mode | |
2838 && GET_MODE (SUBREG_REG (tempreg)) == GET_MODE (data->entry_parm) | |
2839 && GET_MODE_SIZE (GET_MODE (tempreg)) | |
2840 < GET_MODE_SIZE (GET_MODE (data->entry_parm))) | |
2841 { | |
2842 /* The argument is already sign/zero extended, so note it | |
2843 into the subreg. */ | |
2844 SUBREG_PROMOTED_VAR_P (tempreg) = 1; | |
2845 SUBREG_PROMOTED_UNSIGNED_SET (tempreg, unsignedp); | |
2846 } | |
2847 | |
2848 /* TREE_USED gets set erroneously during expand_assignment. */ | |
2849 save_tree_used = TREE_USED (parm); | |
2850 expand_assignment (parm, make_tree (data->nominal_type, tempreg), false); | |
2851 TREE_USED (parm) = save_tree_used; | |
2852 all->first_conversion_insn = get_insns (); | |
2853 all->last_conversion_insn = get_last_insn (); | |
2854 end_sequence (); | |
2855 | |
2856 did_conversion = true; | |
2857 } | |
2858 else | |
2859 emit_move_insn (parmreg, validize_mem (data->entry_parm)); | |
2860 | |
2861 /* If we were passed a pointer but the actual value can safely live | |
2862 in a register, put it in one. */ | |
2863 if (data->passed_pointer | |
2864 && TYPE_MODE (TREE_TYPE (parm)) != BLKmode | |
2865 /* If by-reference argument was promoted, demote it. */ | |
2866 && (TYPE_MODE (TREE_TYPE (parm)) != GET_MODE (DECL_RTL (parm)) | |
2867 || use_register_for_decl (parm))) | |
2868 { | |
2869 /* We can't use nominal_mode, because it will have been set to | |
2870 Pmode above. We must use the actual mode of the parm. */ | |
2871 parmreg = gen_reg_rtx (TYPE_MODE (TREE_TYPE (parm))); | |
2872 mark_user_reg (parmreg); | |
2873 | |
2874 if (GET_MODE (parmreg) != GET_MODE (DECL_RTL (parm))) | |
2875 { | |
2876 rtx tempreg = gen_reg_rtx (GET_MODE (DECL_RTL (parm))); | |
2877 int unsigned_p = TYPE_UNSIGNED (TREE_TYPE (parm)); | |
2878 | |
2879 push_to_sequence2 (all->first_conversion_insn, | |
2880 all->last_conversion_insn); | |
2881 emit_move_insn (tempreg, DECL_RTL (parm)); | |
2882 tempreg = convert_to_mode (GET_MODE (parmreg), tempreg, unsigned_p); | |
2883 emit_move_insn (parmreg, tempreg); | |
2884 all->first_conversion_insn = get_insns (); | |
2885 all->last_conversion_insn = get_last_insn (); | |
2886 end_sequence (); | |
2887 | |
2888 did_conversion = true; | |
2889 } | |
2890 else | |
2891 emit_move_insn (parmreg, DECL_RTL (parm)); | |
2892 | |
2893 SET_DECL_RTL (parm, parmreg); | |
2894 | |
2895 /* STACK_PARM is the pointer, not the parm, and PARMREG is | |
2896 now the parm. */ | |
2897 data->stack_parm = NULL; | |
2898 } | |
2899 | |
2900 /* Mark the register as eliminable if we did no conversion and it was | |
2901 copied from memory at a fixed offset, and the arg pointer was not | |
2902 copied to a pseudo-reg. If the arg pointer is a pseudo reg or the | |
2903 offset formed an invalid address, such memory-equivalences as we | |
2904 make here would screw up life analysis for it. */ | |
2905 if (data->nominal_mode == data->passed_mode | |
2906 && !did_conversion | |
2907 && data->stack_parm != 0 | |
2908 && MEM_P (data->stack_parm) | |
2909 && data->locate.offset.var == 0 | |
2910 && reg_mentioned_p (virtual_incoming_args_rtx, | |
2911 XEXP (data->stack_parm, 0))) | |
2912 { | |
2913 rtx linsn = get_last_insn (); | |
2914 rtx sinsn, set; | |
2915 | |
2916 /* Mark complex types separately. */ | |
2917 if (GET_CODE (parmreg) == CONCAT) | |
2918 { | |
2919 enum machine_mode submode | |
2920 = GET_MODE_INNER (GET_MODE (parmreg)); | |
2921 int regnor = REGNO (XEXP (parmreg, 0)); | |
2922 int regnoi = REGNO (XEXP (parmreg, 1)); | |
2923 rtx stackr = adjust_address_nv (data->stack_parm, submode, 0); | |
2924 rtx stacki = adjust_address_nv (data->stack_parm, submode, | |
2925 GET_MODE_SIZE (submode)); | |
2926 | |
2927 /* Scan backwards for the set of the real and | |
2928 imaginary parts. */ | |
2929 for (sinsn = linsn; sinsn != 0; | |
2930 sinsn = prev_nonnote_insn (sinsn)) | |
2931 { | |
2932 set = single_set (sinsn); | |
2933 if (set == 0) | |
2934 continue; | |
2935 | |
2936 if (SET_DEST (set) == regno_reg_rtx [regnoi]) | |
2937 set_unique_reg_note (sinsn, REG_EQUIV, stacki); | |
2938 else if (SET_DEST (set) == regno_reg_rtx [regnor]) | |
2939 set_unique_reg_note (sinsn, REG_EQUIV, stackr); | |
2940 } | |
2941 } | |
2942 else if ((set = single_set (linsn)) != 0 | |
2943 && SET_DEST (set) == parmreg) | |
2944 set_unique_reg_note (linsn, REG_EQUIV, data->stack_parm); | |
2945 } | |
2946 | |
2947 /* For pointer data type, suggest pointer register. */ | |
2948 if (POINTER_TYPE_P (TREE_TYPE (parm))) | |
2949 mark_reg_pointer (parmreg, | |
2950 TYPE_ALIGN (TREE_TYPE (TREE_TYPE (parm)))); | |
2951 } | |
2952 | |
2953 /* A subroutine of assign_parms. Allocate stack space to hold the current | |
2954 parameter. Get it there. Perform all ABI specified conversions. */ | |
2955 | |
2956 static void | |
2957 assign_parm_setup_stack (struct assign_parm_data_all *all, tree parm, | |
2958 struct assign_parm_data_one *data) | |
2959 { | |
2960 /* Value must be stored in the stack slot STACK_PARM during function | |
2961 execution. */ | |
2962 bool to_conversion = false; | |
2963 | |
2964 assign_parm_remove_parallels (data); | |
2965 | |
2966 if (data->promoted_mode != data->nominal_mode) | |
2967 { | |
2968 /* Conversion is required. */ | |
2969 rtx tempreg = gen_reg_rtx (GET_MODE (data->entry_parm)); | |
2970 | |
2971 emit_move_insn (tempreg, validize_mem (data->entry_parm)); | |
2972 | |
2973 push_to_sequence2 (all->first_conversion_insn, all->last_conversion_insn); | |
2974 to_conversion = true; | |
2975 | |
2976 data->entry_parm = convert_to_mode (data->nominal_mode, tempreg, | |
2977 TYPE_UNSIGNED (TREE_TYPE (parm))); | |
2978 | |
2979 if (data->stack_parm) | |
2980 /* ??? This may need a big-endian conversion on sparc64. */ | |
2981 data->stack_parm | |
2982 = adjust_address (data->stack_parm, data->nominal_mode, 0); | |
2983 } | |
2984 | |
2985 if (data->entry_parm != data->stack_parm) | |
2986 { | |
2987 rtx src, dest; | |
2988 | |
2989 if (data->stack_parm == 0) | |
2990 { | |
2991 int align = STACK_SLOT_ALIGNMENT (data->passed_type, | |
2992 GET_MODE (data->entry_parm), | |
2993 TYPE_ALIGN (data->passed_type)); | |
2994 data->stack_parm | |
2995 = assign_stack_local (GET_MODE (data->entry_parm), | |
2996 GET_MODE_SIZE (GET_MODE (data->entry_parm)), | |
2997 align); | |
2998 set_mem_attributes (data->stack_parm, parm, 1); | |
2999 } | |
3000 | |
3001 dest = validize_mem (data->stack_parm); | |
3002 src = validize_mem (data->entry_parm); | |
3003 | |
3004 if (MEM_P (src)) | |
3005 { | |
3006 /* Use a block move to handle potentially misaligned entry_parm. */ | |
3007 if (!to_conversion) | |
3008 push_to_sequence2 (all->first_conversion_insn, | |
3009 all->last_conversion_insn); | |
3010 to_conversion = true; | |
3011 | |
3012 emit_block_move (dest, src, | |
3013 GEN_INT (int_size_in_bytes (data->passed_type)), | |
3014 BLOCK_OP_NORMAL); | |
3015 } | |
3016 else | |
3017 emit_move_insn (dest, src); | |
3018 } | |
3019 | |
3020 if (to_conversion) | |
3021 { | |
3022 all->first_conversion_insn = get_insns (); | |
3023 all->last_conversion_insn = get_last_insn (); | |
3024 end_sequence (); | |
3025 } | |
3026 | |
3027 SET_DECL_RTL (parm, data->stack_parm); | |
3028 } | |
3029 | |
3030 /* A subroutine of assign_parms. If the ABI splits complex arguments, then | |
3031 undo the frobbing that we did in assign_parms_augmented_arg_list. */ | |
3032 | |
3033 static void | |
3034 assign_parms_unsplit_complex (struct assign_parm_data_all *all, tree fnargs) | |
3035 { | |
3036 tree parm; | |
3037 tree orig_fnargs = all->orig_fnargs; | |
3038 | |
3039 for (parm = orig_fnargs; parm; parm = TREE_CHAIN (parm)) | |
3040 { | |
3041 if (TREE_CODE (TREE_TYPE (parm)) == COMPLEX_TYPE | |
3042 && targetm.calls.split_complex_arg (TREE_TYPE (parm))) | |
3043 { | |
3044 rtx tmp, real, imag; | |
3045 enum machine_mode inner = GET_MODE_INNER (DECL_MODE (parm)); | |
3046 | |
3047 real = DECL_RTL (fnargs); | |
3048 imag = DECL_RTL (TREE_CHAIN (fnargs)); | |
3049 if (inner != GET_MODE (real)) | |
3050 { | |
3051 real = gen_lowpart_SUBREG (inner, real); | |
3052 imag = gen_lowpart_SUBREG (inner, imag); | |
3053 } | |
3054 | |
3055 if (TREE_ADDRESSABLE (parm)) | |
3056 { | |
3057 rtx rmem, imem; | |
3058 HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (parm)); | |
3059 int align = STACK_SLOT_ALIGNMENT (TREE_TYPE (parm), | |
3060 DECL_MODE (parm), | |
3061 TYPE_ALIGN (TREE_TYPE (parm))); | |
3062 | |
3063 /* split_complex_arg put the real and imag parts in | |
3064 pseudos. Move them to memory. */ | |
3065 tmp = assign_stack_local (DECL_MODE (parm), size, align); | |
3066 set_mem_attributes (tmp, parm, 1); | |
3067 rmem = adjust_address_nv (tmp, inner, 0); | |
3068 imem = adjust_address_nv (tmp, inner, GET_MODE_SIZE (inner)); | |
3069 push_to_sequence2 (all->first_conversion_insn, | |
3070 all->last_conversion_insn); | |
3071 emit_move_insn (rmem, real); | |
3072 emit_move_insn (imem, imag); | |
3073 all->first_conversion_insn = get_insns (); | |
3074 all->last_conversion_insn = get_last_insn (); | |
3075 end_sequence (); | |
3076 } | |
3077 else | |
3078 tmp = gen_rtx_CONCAT (DECL_MODE (parm), real, imag); | |
3079 SET_DECL_RTL (parm, tmp); | |
3080 | |
3081 real = DECL_INCOMING_RTL (fnargs); | |
3082 imag = DECL_INCOMING_RTL (TREE_CHAIN (fnargs)); | |
3083 if (inner != GET_MODE (real)) | |
3084 { | |
3085 real = gen_lowpart_SUBREG (inner, real); | |
3086 imag = gen_lowpart_SUBREG (inner, imag); | |
3087 } | |
3088 tmp = gen_rtx_CONCAT (DECL_MODE (parm), real, imag); | |
3089 set_decl_incoming_rtl (parm, tmp, false); | |
3090 fnargs = TREE_CHAIN (fnargs); | |
3091 } | |
3092 else | |
3093 { | |
3094 SET_DECL_RTL (parm, DECL_RTL (fnargs)); | |
3095 set_decl_incoming_rtl (parm, DECL_INCOMING_RTL (fnargs), false); | |
3096 | |
3097 /* Set MEM_EXPR to the original decl, i.e. to PARM, | |
3098 instead of the copy of decl, i.e. FNARGS. */ | |
3099 if (DECL_INCOMING_RTL (parm) && MEM_P (DECL_INCOMING_RTL (parm))) | |
3100 set_mem_expr (DECL_INCOMING_RTL (parm), parm); | |
3101 } | |
3102 | |
3103 fnargs = TREE_CHAIN (fnargs); | |
3104 } | |
3105 } | |
3106 | |
3107 /* Assign RTL expressions to the function's parameters. This may involve | |
3108 copying them into registers and using those registers as the DECL_RTL. */ | |
3109 | |
3110 static void | |
3111 assign_parms (tree fndecl) | |
3112 { | |
3113 struct assign_parm_data_all all; | |
3114 tree fnargs, parm; | |
3115 | |
3116 crtl->args.internal_arg_pointer | |
3117 = targetm.calls.internal_arg_pointer (); | |
3118 | |
3119 assign_parms_initialize_all (&all); | |
3120 fnargs = assign_parms_augmented_arg_list (&all); | |
3121 | |
3122 for (parm = fnargs; parm; parm = TREE_CHAIN (parm)) | |
3123 { | |
3124 struct assign_parm_data_one data; | |
3125 | |
3126 /* Extract the type of PARM; adjust it according to ABI. */ | |
3127 assign_parm_find_data_types (&all, parm, &data); | |
3128 | |
3129 /* Early out for errors and void parameters. */ | |
3130 if (data.passed_mode == VOIDmode) | |
3131 { | |
3132 SET_DECL_RTL (parm, const0_rtx); | |
3133 DECL_INCOMING_RTL (parm) = DECL_RTL (parm); | |
3134 continue; | |
3135 } | |
3136 | |
3137 /* Estimate stack alignment from parameter alignment. */ | |
3138 if (SUPPORTS_STACK_ALIGNMENT) | |
3139 { | |
3140 unsigned int align = FUNCTION_ARG_BOUNDARY (data.promoted_mode, | |
3141 data.passed_type); | |
3142 if (TYPE_ALIGN (data.nominal_type) > align) | |
3143 align = TYPE_ALIGN (data.passed_type); | |
3144 if (crtl->stack_alignment_estimated < align) | |
3145 { | |
3146 gcc_assert (!crtl->stack_realign_processed); | |
3147 crtl->stack_alignment_estimated = align; | |
3148 } | |
3149 } | |
3150 | |
3151 if (cfun->stdarg && !TREE_CHAIN (parm)) | |
3152 assign_parms_setup_varargs (&all, &data, false); | |
3153 | |
3154 /* Find out where the parameter arrives in this function. */ | |
3155 assign_parm_find_entry_rtl (&all, &data); | |
3156 | |
3157 /* Find out where stack space for this parameter might be. */ | |
3158 if (assign_parm_is_stack_parm (&all, &data)) | |
3159 { | |
3160 assign_parm_find_stack_rtl (parm, &data); | |
3161 assign_parm_adjust_entry_rtl (&data); | |
3162 } | |
3163 | |
3164 /* Record permanently how this parm was passed. */ | |
3165 set_decl_incoming_rtl (parm, data.entry_parm, data.passed_pointer); | |
3166 | |
3167 /* Update info on where next arg arrives in registers. */ | |
3168 FUNCTION_ARG_ADVANCE (all.args_so_far, data.promoted_mode, | |
3169 data.passed_type, data.named_arg); | |
3170 | |
3171 assign_parm_adjust_stack_rtl (&data); | |
3172 | |
3173 if (assign_parm_setup_block_p (&data)) | |
3174 assign_parm_setup_block (&all, parm, &data); | |
3175 else if (data.passed_pointer || use_register_for_decl (parm)) | |
3176 assign_parm_setup_reg (&all, parm, &data); | |
3177 else | |
3178 assign_parm_setup_stack (&all, parm, &data); | |
3179 } | |
3180 | |
3181 if (targetm.calls.split_complex_arg && fnargs != all.orig_fnargs) | |
3182 assign_parms_unsplit_complex (&all, fnargs); | |
3183 | |
3184 /* Output all parameter conversion instructions (possibly including calls) | |
3185 now that all parameters have been copied out of hard registers. */ | |
3186 emit_insn (all.first_conversion_insn); | |
3187 | |
3188 /* Estimate reload stack alignment from scalar return mode. */ | |
3189 if (SUPPORTS_STACK_ALIGNMENT) | |
3190 { | |
3191 if (DECL_RESULT (fndecl)) | |
3192 { | |
3193 tree type = TREE_TYPE (DECL_RESULT (fndecl)); | |
3194 enum machine_mode mode = TYPE_MODE (type); | |
3195 | |
3196 if (mode != BLKmode | |
3197 && mode != VOIDmode | |
3198 && !AGGREGATE_TYPE_P (type)) | |
3199 { | |
3200 unsigned int align = GET_MODE_ALIGNMENT (mode); | |
3201 if (crtl->stack_alignment_estimated < align) | |
3202 { | |
3203 gcc_assert (!crtl->stack_realign_processed); | |
3204 crtl->stack_alignment_estimated = align; | |
3205 } | |
3206 } | |
3207 } | |
3208 } | |
3209 | |
3210 /* If we are receiving a struct value address as the first argument, set up | |
3211 the RTL for the function result. As this might require code to convert | |
3212 the transmitted address to Pmode, we do this here to ensure that possible | |
3213 preliminary conversions of the address have been emitted already. */ | |
3214 if (all.function_result_decl) | |
3215 { | |
3216 tree result = DECL_RESULT (current_function_decl); | |
3217 rtx addr = DECL_RTL (all.function_result_decl); | |
3218 rtx x; | |
3219 | |
3220 if (DECL_BY_REFERENCE (result)) | |
3221 x = addr; | |
3222 else | |
3223 { | |
3224 addr = convert_memory_address (Pmode, addr); | |
3225 x = gen_rtx_MEM (DECL_MODE (result), addr); | |
3226 set_mem_attributes (x, result, 1); | |
3227 } | |
3228 SET_DECL_RTL (result, x); | |
3229 } | |
3230 | |
3231 /* We have aligned all the args, so add space for the pretend args. */ | |
3232 crtl->args.pretend_args_size = all.pretend_args_size; | |
3233 all.stack_args_size.constant += all.extra_pretend_bytes; | |
3234 crtl->args.size = all.stack_args_size.constant; | |
3235 | |
3236 /* Adjust function incoming argument size for alignment and | |
3237 minimum length. */ | |
3238 | |
3239 #ifdef REG_PARM_STACK_SPACE | |
3240 crtl->args.size = MAX (crtl->args.size, | |
3241 REG_PARM_STACK_SPACE (fndecl)); | |
3242 #endif | |
3243 | |
3244 crtl->args.size = CEIL_ROUND (crtl->args.size, | |
3245 PARM_BOUNDARY / BITS_PER_UNIT); | |
3246 | |
3247 #ifdef ARGS_GROW_DOWNWARD | |
3248 crtl->args.arg_offset_rtx | |
3249 = (all.stack_args_size.var == 0 ? GEN_INT (-all.stack_args_size.constant) | |
3250 : expand_expr (size_diffop (all.stack_args_size.var, | |
3251 size_int (-all.stack_args_size.constant)), | |
3252 NULL_RTX, VOIDmode, 0)); | |
3253 #else | |
3254 crtl->args.arg_offset_rtx = ARGS_SIZE_RTX (all.stack_args_size); | |
3255 #endif | |
3256 | |
3257 /* See how many bytes, if any, of its args a function should try to pop | |
3258 on return. */ | |
3259 | |
3260 crtl->args.pops_args = RETURN_POPS_ARGS (fndecl, TREE_TYPE (fndecl), | |
3261 crtl->args.size); | |
3262 | |
3263 /* For stdarg.h function, save info about | |
3264 regs and stack space used by the named args. */ | |
3265 | |
3266 crtl->args.info = all.args_so_far; | |
3267 | |
3268 /* Set the rtx used for the function return value. Put this in its | |
3269 own variable so any optimizers that need this information don't have | |
3270 to include tree.h. Do this here so it gets done when an inlined | |
3271 function gets output. */ | |
3272 | |
3273 crtl->return_rtx | |
3274 = (DECL_RTL_SET_P (DECL_RESULT (fndecl)) | |
3275 ? DECL_RTL (DECL_RESULT (fndecl)) : NULL_RTX); | |
3276 | |
3277 /* If scalar return value was computed in a pseudo-reg, or was a named | |
3278 return value that got dumped to the stack, copy that to the hard | |
3279 return register. */ | |
3280 if (DECL_RTL_SET_P (DECL_RESULT (fndecl))) | |
3281 { | |
3282 tree decl_result = DECL_RESULT (fndecl); | |
3283 rtx decl_rtl = DECL_RTL (decl_result); | |
3284 | |
3285 if (REG_P (decl_rtl) | |
3286 ? REGNO (decl_rtl) >= FIRST_PSEUDO_REGISTER | |
3287 : DECL_REGISTER (decl_result)) | |
3288 { | |
3289 rtx real_decl_rtl; | |
3290 | |
3291 real_decl_rtl = targetm.calls.function_value (TREE_TYPE (decl_result), | |
3292 fndecl, true); | |
3293 REG_FUNCTION_VALUE_P (real_decl_rtl) = 1; | |
3294 /* The delay slot scheduler assumes that crtl->return_rtx | |
3295 holds the hard register containing the return value, not a | |
3296 temporary pseudo. */ | |
3297 crtl->return_rtx = real_decl_rtl; | |
3298 } | |
3299 } | |
3300 } | |
3301 | |
3302 /* A subroutine of gimplify_parameters, invoked via walk_tree. | |
3303 For all seen types, gimplify their sizes. */ | |
3304 | |
3305 static tree | |
3306 gimplify_parm_type (tree *tp, int *walk_subtrees, void *data) | |
3307 { | |
3308 tree t = *tp; | |
3309 | |
3310 *walk_subtrees = 0; | |
3311 if (TYPE_P (t)) | |
3312 { | |
3313 if (POINTER_TYPE_P (t)) | |
3314 *walk_subtrees = 1; | |
3315 else if (TYPE_SIZE (t) && !TREE_CONSTANT (TYPE_SIZE (t)) | |
3316 && !TYPE_SIZES_GIMPLIFIED (t)) | |
3317 { | |
3318 gimplify_type_sizes (t, (gimple_seq *) data); | |
3319 *walk_subtrees = 1; | |
3320 } | |
3321 } | |
3322 | |
3323 return NULL; | |
3324 } | |
3325 | |
3326 /* Gimplify the parameter list for current_function_decl. This involves | |
3327 evaluating SAVE_EXPRs of variable sized parameters and generating code | |
3328 to implement callee-copies reference parameters. Returns a sequence of | |
3329 statements to add to the beginning of the function. */ | |
3330 | |
3331 gimple_seq | |
3332 gimplify_parameters (void) | |
3333 { | |
3334 struct assign_parm_data_all all; | |
3335 tree fnargs, parm; | |
3336 gimple_seq stmts = NULL; | |
3337 | |
3338 assign_parms_initialize_all (&all); | |
3339 fnargs = assign_parms_augmented_arg_list (&all); | |
3340 | |
3341 for (parm = fnargs; parm; parm = TREE_CHAIN (parm)) | |
3342 { | |
3343 struct assign_parm_data_one data; | |
3344 | |
3345 /* Extract the type of PARM; adjust it according to ABI. */ | |
3346 assign_parm_find_data_types (&all, parm, &data); | |
3347 | |
3348 /* Early out for errors and void parameters. */ | |
3349 if (data.passed_mode == VOIDmode || DECL_SIZE (parm) == NULL) | |
3350 continue; | |
3351 | |
3352 /* Update info on where next arg arrives in registers. */ | |
3353 FUNCTION_ARG_ADVANCE (all.args_so_far, data.promoted_mode, | |
3354 data.passed_type, data.named_arg); | |
3355 | |
3356 /* ??? Once upon a time variable_size stuffed parameter list | |
3357 SAVE_EXPRs (amongst others) onto a pending sizes list. This | |
3358 turned out to be less than manageable in the gimple world. | |
3359 Now we have to hunt them down ourselves. */ | |
3360 walk_tree_without_duplicates (&data.passed_type, | |
3361 gimplify_parm_type, &stmts); | |
3362 | |
3363 if (TREE_CODE (DECL_SIZE_UNIT (parm)) != INTEGER_CST) | |
3364 { | |
3365 gimplify_one_sizepos (&DECL_SIZE (parm), &stmts); | |
3366 gimplify_one_sizepos (&DECL_SIZE_UNIT (parm), &stmts); | |
3367 } | |
3368 | |
3369 if (data.passed_pointer) | |
3370 { | |
3371 tree type = TREE_TYPE (data.passed_type); | |
3372 if (reference_callee_copied (&all.args_so_far, TYPE_MODE (type), | |
3373 type, data.named_arg)) | |
3374 { | |
3375 tree local, t; | |
3376 | |
3377 /* For constant-sized objects, this is trivial; for | |
3378 variable-sized objects, we have to play games. */ | |
3379 if (TREE_CODE (DECL_SIZE_UNIT (parm)) == INTEGER_CST | |
3380 && !(flag_stack_check == GENERIC_STACK_CHECK | |
3381 && compare_tree_int (DECL_SIZE_UNIT (parm), | |
3382 STACK_CHECK_MAX_VAR_SIZE) > 0)) | |
3383 { | |
3384 local = create_tmp_var (type, get_name (parm)); | |
3385 DECL_IGNORED_P (local) = 0; | |
3386 /* If PARM was addressable, move that flag over | |
3387 to the local copy, as its address will be taken, | |
3388 not the PARMs. */ | |
3389 if (TREE_ADDRESSABLE (parm)) | |
3390 { | |
3391 TREE_ADDRESSABLE (parm) = 0; | |
3392 TREE_ADDRESSABLE (local) = 1; | |
3393 } | |
3394 } | |
3395 else | |
3396 { | |
3397 tree ptr_type, addr; | |
3398 | |
3399 ptr_type = build_pointer_type (type); | |
3400 addr = create_tmp_var (ptr_type, get_name (parm)); | |
3401 DECL_IGNORED_P (addr) = 0; | |
3402 local = build_fold_indirect_ref (addr); | |
3403 | |
3404 t = built_in_decls[BUILT_IN_ALLOCA]; | |
3405 t = build_call_expr (t, 1, DECL_SIZE_UNIT (parm)); | |
3406 t = fold_convert (ptr_type, t); | |
3407 t = build2 (MODIFY_EXPR, TREE_TYPE (addr), addr, t); | |
3408 gimplify_and_add (t, &stmts); | |
3409 } | |
3410 | |
3411 gimplify_assign (local, parm, &stmts); | |
3412 | |
3413 SET_DECL_VALUE_EXPR (parm, local); | |
3414 DECL_HAS_VALUE_EXPR_P (parm) = 1; | |
3415 } | |
3416 } | |
3417 } | |
3418 | |
3419 return stmts; | |
3420 } | |
3421 | |
3422 /* Compute the size and offset from the start of the stacked arguments for a | |
3423 parm passed in mode PASSED_MODE and with type TYPE. | |
3424 | |
3425 INITIAL_OFFSET_PTR points to the current offset into the stacked | |
3426 arguments. | |
3427 | |
3428 The starting offset and size for this parm are returned in | |
3429 LOCATE->OFFSET and LOCATE->SIZE, respectively. When IN_REGS is | |
3430 nonzero, the offset is that of stack slot, which is returned in | |
3431 LOCATE->SLOT_OFFSET. LOCATE->ALIGNMENT_PAD is the amount of | |
3432 padding required from the initial offset ptr to the stack slot. | |
3433 | |
3434 IN_REGS is nonzero if the argument will be passed in registers. It will | |
3435 never be set if REG_PARM_STACK_SPACE is not defined. | |
3436 | |
3437 FNDECL is the function in which the argument was defined. | |
3438 | |
3439 There are two types of rounding that are done. The first, controlled by | |
3440 FUNCTION_ARG_BOUNDARY, forces the offset from the start of the argument | |
3441 list to be aligned to the specific boundary (in bits). This rounding | |
3442 affects the initial and starting offsets, but not the argument size. | |
3443 | |
3444 The second, controlled by FUNCTION_ARG_PADDING and PARM_BOUNDARY, | |
3445 optionally rounds the size of the parm to PARM_BOUNDARY. The | |
3446 initial offset is not affected by this rounding, while the size always | |
3447 is and the starting offset may be. */ | |
3448 | |
3449 /* LOCATE->OFFSET will be negative for ARGS_GROW_DOWNWARD case; | |
3450 INITIAL_OFFSET_PTR is positive because locate_and_pad_parm's | |
3451 callers pass in the total size of args so far as | |
3452 INITIAL_OFFSET_PTR. LOCATE->SIZE is always positive. */ | |
3453 | |
3454 void | |
3455 locate_and_pad_parm (enum machine_mode passed_mode, tree type, int in_regs, | |
3456 int partial, tree fndecl ATTRIBUTE_UNUSED, | |
3457 struct args_size *initial_offset_ptr, | |
3458 struct locate_and_pad_arg_data *locate) | |
3459 { | |
3460 tree sizetree; | |
3461 enum direction where_pad; | |
3462 unsigned int boundary; | |
3463 int reg_parm_stack_space = 0; | |
3464 int part_size_in_regs; | |
3465 | |
3466 #ifdef REG_PARM_STACK_SPACE | |
3467 reg_parm_stack_space = REG_PARM_STACK_SPACE (fndecl); | |
3468 | |
3469 /* If we have found a stack parm before we reach the end of the | |
3470 area reserved for registers, skip that area. */ | |
3471 if (! in_regs) | |
3472 { | |
3473 if (reg_parm_stack_space > 0) | |
3474 { | |
3475 if (initial_offset_ptr->var) | |
3476 { | |
3477 initial_offset_ptr->var | |
3478 = size_binop (MAX_EXPR, ARGS_SIZE_TREE (*initial_offset_ptr), | |
3479 ssize_int (reg_parm_stack_space)); | |
3480 initial_offset_ptr->constant = 0; | |
3481 } | |
3482 else if (initial_offset_ptr->constant < reg_parm_stack_space) | |
3483 initial_offset_ptr->constant = reg_parm_stack_space; | |
3484 } | |
3485 } | |
3486 #endif /* REG_PARM_STACK_SPACE */ | |
3487 | |
3488 part_size_in_regs = (reg_parm_stack_space == 0 ? partial : 0); | |
3489 | |
3490 sizetree | |
3491 = type ? size_in_bytes (type) : size_int (GET_MODE_SIZE (passed_mode)); | |
3492 where_pad = FUNCTION_ARG_PADDING (passed_mode, type); | |
3493 boundary = FUNCTION_ARG_BOUNDARY (passed_mode, type); | |
3494 locate->where_pad = where_pad; | |
3495 | |
3496 /* Alignment can't exceed MAX_SUPPORTED_STACK_ALIGNMENT. */ | |
3497 if (boundary > MAX_SUPPORTED_STACK_ALIGNMENT) | |
3498 boundary = MAX_SUPPORTED_STACK_ALIGNMENT; | |
3499 | |
3500 locate->boundary = boundary; | |
3501 | |
3502 if (SUPPORTS_STACK_ALIGNMENT) | |
3503 { | |
3504 /* stack_alignment_estimated can't change after stack has been | |
3505 realigned. */ | |
3506 if (crtl->stack_alignment_estimated < boundary) | |
3507 { | |
3508 if (!crtl->stack_realign_processed) | |
3509 crtl->stack_alignment_estimated = boundary; | |
3510 else | |
3511 { | |
3512 /* If stack is realigned and stack alignment value | |
3513 hasn't been finalized, it is OK not to increase | |
3514 stack_alignment_estimated. The bigger alignment | |
3515 requirement is recorded in stack_alignment_needed | |
3516 below. */ | |
3517 gcc_assert (!crtl->stack_realign_finalized | |
3518 && crtl->stack_realign_needed); | |
3519 } | |
3520 } | |
3521 } | |
3522 | |
3523 /* Remember if the outgoing parameter requires extra alignment on the | |
3524 calling function side. */ | |
3525 if (crtl->stack_alignment_needed < boundary) | |
3526 crtl->stack_alignment_needed = boundary; | |
3527 if (crtl->max_used_stack_slot_alignment < crtl->stack_alignment_needed) | |
3528 crtl->max_used_stack_slot_alignment = crtl->stack_alignment_needed; | |
3529 if (crtl->preferred_stack_boundary < boundary) | |
3530 crtl->preferred_stack_boundary = boundary; | |
3531 | |
3532 #ifdef ARGS_GROW_DOWNWARD | |
3533 locate->slot_offset.constant = -initial_offset_ptr->constant; | |
3534 if (initial_offset_ptr->var) | |
3535 locate->slot_offset.var = size_binop (MINUS_EXPR, ssize_int (0), | |
3536 initial_offset_ptr->var); | |
3537 | |
3538 { | |
3539 tree s2 = sizetree; | |
3540 if (where_pad != none | |
3541 && (!host_integerp (sizetree, 1) | |
3542 || (tree_low_cst (sizetree, 1) * BITS_PER_UNIT) % PARM_BOUNDARY)) | |
3543 s2 = round_up (s2, PARM_BOUNDARY / BITS_PER_UNIT); | |
3544 SUB_PARM_SIZE (locate->slot_offset, s2); | |
3545 } | |
3546 | |
3547 locate->slot_offset.constant += part_size_in_regs; | |
3548 | |
3549 if (!in_regs | |
3550 #ifdef REG_PARM_STACK_SPACE | |
3551 || REG_PARM_STACK_SPACE (fndecl) > 0 | |
3552 #endif | |
3553 ) | |
3554 pad_to_arg_alignment (&locate->slot_offset, boundary, | |
3555 &locate->alignment_pad); | |
3556 | |
3557 locate->size.constant = (-initial_offset_ptr->constant | |
3558 - locate->slot_offset.constant); | |
3559 if (initial_offset_ptr->var) | |
3560 locate->size.var = size_binop (MINUS_EXPR, | |
3561 size_binop (MINUS_EXPR, | |
3562 ssize_int (0), | |
3563 initial_offset_ptr->var), | |
3564 locate->slot_offset.var); | |
3565 | |
3566 /* Pad_below needs the pre-rounded size to know how much to pad | |
3567 below. */ | |
3568 locate->offset = locate->slot_offset; | |
3569 if (where_pad == downward) | |
3570 pad_below (&locate->offset, passed_mode, sizetree); | |
3571 | |
3572 #else /* !ARGS_GROW_DOWNWARD */ | |
3573 if (!in_regs | |
3574 #ifdef REG_PARM_STACK_SPACE | |
3575 || REG_PARM_STACK_SPACE (fndecl) > 0 | |
3576 #endif | |
3577 ) | |
3578 pad_to_arg_alignment (initial_offset_ptr, boundary, | |
3579 &locate->alignment_pad); | |
3580 locate->slot_offset = *initial_offset_ptr; | |
3581 | |
3582 #ifdef PUSH_ROUNDING | |
3583 if (passed_mode != BLKmode) | |
3584 sizetree = size_int (PUSH_ROUNDING (TREE_INT_CST_LOW (sizetree))); | |
3585 #endif | |
3586 | |
3587 /* Pad_below needs the pre-rounded size to know how much to pad below | |
3588 so this must be done before rounding up. */ | |
3589 locate->offset = locate->slot_offset; | |
3590 if (where_pad == downward) | |
3591 pad_below (&locate->offset, passed_mode, sizetree); | |
3592 | |
3593 if (where_pad != none | |
3594 && (!host_integerp (sizetree, 1) | |
3595 || (tree_low_cst (sizetree, 1) * BITS_PER_UNIT) % PARM_BOUNDARY)) | |
3596 sizetree = round_up (sizetree, PARM_BOUNDARY / BITS_PER_UNIT); | |
3597 | |
3598 ADD_PARM_SIZE (locate->size, sizetree); | |
3599 | |
3600 locate->size.constant -= part_size_in_regs; | |
3601 #endif /* ARGS_GROW_DOWNWARD */ | |
3602 | |
3603 #ifdef FUNCTION_ARG_OFFSET | |
3604 locate->offset.constant += FUNCTION_ARG_OFFSET (passed_mode, type); | |
3605 #endif | |
3606 } | |
3607 | |
3608 /* Round the stack offset in *OFFSET_PTR up to a multiple of BOUNDARY. | |
3609 BOUNDARY is measured in bits, but must be a multiple of a storage unit. */ | |
3610 | |
3611 static void | |
3612 pad_to_arg_alignment (struct args_size *offset_ptr, int boundary, | |
3613 struct args_size *alignment_pad) | |
3614 { | |
3615 tree save_var = NULL_TREE; | |
3616 HOST_WIDE_INT save_constant = 0; | |
3617 int boundary_in_bytes = boundary / BITS_PER_UNIT; | |
3618 HOST_WIDE_INT sp_offset = STACK_POINTER_OFFSET; | |
3619 | |
3620 #ifdef SPARC_STACK_BOUNDARY_HACK | |
3621 /* ??? The SPARC port may claim a STACK_BOUNDARY higher than | |
3622 the real alignment of %sp. However, when it does this, the | |
3623 alignment of %sp+STACK_POINTER_OFFSET is STACK_BOUNDARY. */ | |
3624 if (SPARC_STACK_BOUNDARY_HACK) | |
3625 sp_offset = 0; | |
3626 #endif | |
3627 | |
3628 if (boundary > PARM_BOUNDARY) | |
3629 { | |
3630 save_var = offset_ptr->var; | |
3631 save_constant = offset_ptr->constant; | |
3632 } | |
3633 | |
3634 alignment_pad->var = NULL_TREE; | |
3635 alignment_pad->constant = 0; | |
3636 | |
3637 if (boundary > BITS_PER_UNIT) | |
3638 { | |
3639 if (offset_ptr->var) | |
3640 { | |
3641 tree sp_offset_tree = ssize_int (sp_offset); | |
3642 tree offset = size_binop (PLUS_EXPR, | |
3643 ARGS_SIZE_TREE (*offset_ptr), | |
3644 sp_offset_tree); | |
3645 #ifdef ARGS_GROW_DOWNWARD | |
3646 tree rounded = round_down (offset, boundary / BITS_PER_UNIT); | |
3647 #else | |
3648 tree rounded = round_up (offset, boundary / BITS_PER_UNIT); | |
3649 #endif | |
3650 | |
3651 offset_ptr->var = size_binop (MINUS_EXPR, rounded, sp_offset_tree); | |
3652 /* ARGS_SIZE_TREE includes constant term. */ | |
3653 offset_ptr->constant = 0; | |
3654 if (boundary > PARM_BOUNDARY) | |
3655 alignment_pad->var = size_binop (MINUS_EXPR, offset_ptr->var, | |
3656 save_var); | |
3657 } | |
3658 else | |
3659 { | |
3660 offset_ptr->constant = -sp_offset + | |
3661 #ifdef ARGS_GROW_DOWNWARD | |
3662 FLOOR_ROUND (offset_ptr->constant + sp_offset, boundary_in_bytes); | |
3663 #else | |
3664 CEIL_ROUND (offset_ptr->constant + sp_offset, boundary_in_bytes); | |
3665 #endif | |
3666 if (boundary > PARM_BOUNDARY) | |
3667 alignment_pad->constant = offset_ptr->constant - save_constant; | |
3668 } | |
3669 } | |
3670 } | |
3671 | |
3672 static void | |
3673 pad_below (struct args_size *offset_ptr, enum machine_mode passed_mode, tree sizetree) | |
3674 { | |
3675 if (passed_mode != BLKmode) | |
3676 { | |
3677 if (GET_MODE_BITSIZE (passed_mode) % PARM_BOUNDARY) | |
3678 offset_ptr->constant | |
3679 += (((GET_MODE_BITSIZE (passed_mode) + PARM_BOUNDARY - 1) | |
3680 / PARM_BOUNDARY * PARM_BOUNDARY / BITS_PER_UNIT) | |
3681 - GET_MODE_SIZE (passed_mode)); | |
3682 } | |
3683 else | |
3684 { | |
3685 if (TREE_CODE (sizetree) != INTEGER_CST | |
3686 || (TREE_INT_CST_LOW (sizetree) * BITS_PER_UNIT) % PARM_BOUNDARY) | |
3687 { | |
3688 /* Round the size up to multiple of PARM_BOUNDARY bits. */ | |
3689 tree s2 = round_up (sizetree, PARM_BOUNDARY / BITS_PER_UNIT); | |
3690 /* Add it in. */ | |
3691 ADD_PARM_SIZE (*offset_ptr, s2); | |
3692 SUB_PARM_SIZE (*offset_ptr, sizetree); | |
3693 } | |
3694 } | |
3695 } | |
3696 | |
3697 | |
3698 /* True if register REGNO was alive at a place where `setjmp' was | |
3699 called and was set more than once or is an argument. Such regs may | |
3700 be clobbered by `longjmp'. */ | |
3701 | |
3702 static bool | |
3703 regno_clobbered_at_setjmp (bitmap setjmp_crosses, int regno) | |
3704 { | |
3705 /* There appear to be cases where some local vars never reach the | |
3706 backend but have bogus regnos. */ | |
3707 if (regno >= max_reg_num ()) | |
3708 return false; | |
3709 | |
3710 return ((REG_N_SETS (regno) > 1 | |
3711 || REGNO_REG_SET_P (df_get_live_out (ENTRY_BLOCK_PTR), regno)) | |
3712 && REGNO_REG_SET_P (setjmp_crosses, regno)); | |
3713 } | |
3714 | |
3715 /* Walk the tree of blocks describing the binding levels within a | |
3716 function and warn about variables the might be killed by setjmp or | |
3717 vfork. This is done after calling flow_analysis before register | |
3718 allocation since that will clobber the pseudo-regs to hard | |
3719 regs. */ | |
3720 | |
3721 static void | |
3722 setjmp_vars_warning (bitmap setjmp_crosses, tree block) | |
3723 { | |
3724 tree decl, sub; | |
3725 | |
3726 for (decl = BLOCK_VARS (block); decl; decl = TREE_CHAIN (decl)) | |
3727 { | |
3728 if (TREE_CODE (decl) == VAR_DECL | |
3729 && DECL_RTL_SET_P (decl) | |
3730 && REG_P (DECL_RTL (decl)) | |
3731 && regno_clobbered_at_setjmp (setjmp_crosses, REGNO (DECL_RTL (decl)))) | |
3732 warning (OPT_Wclobbered, "variable %q+D might be clobbered by" | |
3733 " %<longjmp%> or %<vfork%>", decl); | |
3734 } | |
3735 | |
3736 for (sub = BLOCK_SUBBLOCKS (block); sub; sub = BLOCK_CHAIN (sub)) | |
3737 setjmp_vars_warning (setjmp_crosses, sub); | |
3738 } | |
3739 | |
3740 /* Do the appropriate part of setjmp_vars_warning | |
3741 but for arguments instead of local variables. */ | |
3742 | |
3743 static void | |
3744 setjmp_args_warning (bitmap setjmp_crosses) | |
3745 { | |
3746 tree decl; | |
3747 for (decl = DECL_ARGUMENTS (current_function_decl); | |
3748 decl; decl = TREE_CHAIN (decl)) | |
3749 if (DECL_RTL (decl) != 0 | |
3750 && REG_P (DECL_RTL (decl)) | |
3751 && regno_clobbered_at_setjmp (setjmp_crosses, REGNO (DECL_RTL (decl)))) | |
3752 warning (OPT_Wclobbered, | |
3753 "argument %q+D might be clobbered by %<longjmp%> or %<vfork%>", | |
3754 decl); | |
3755 } | |
3756 | |
3757 /* Generate warning messages for variables live across setjmp. */ | |
3758 | |
3759 void | |
3760 generate_setjmp_warnings (void) | |
3761 { | |
3762 bitmap setjmp_crosses = regstat_get_setjmp_crosses (); | |
3763 | |
3764 if (n_basic_blocks == NUM_FIXED_BLOCKS | |
3765 || bitmap_empty_p (setjmp_crosses)) | |
3766 return; | |
3767 | |
3768 setjmp_vars_warning (setjmp_crosses, DECL_INITIAL (current_function_decl)); | |
3769 setjmp_args_warning (setjmp_crosses); | |
3770 } | |
3771 | |
3772 | |
3773 /* Identify BLOCKs referenced by more than one NOTE_INSN_BLOCK_{BEG,END}, | |
3774 and create duplicate blocks. */ | |
3775 /* ??? Need an option to either create block fragments or to create | |
3776 abstract origin duplicates of a source block. It really depends | |
3777 on what optimization has been performed. */ | |
3778 | |
3779 void | |
3780 reorder_blocks (void) | |
3781 { | |
3782 tree block = DECL_INITIAL (current_function_decl); | |
3783 VEC(tree,heap) *block_stack; | |
3784 | |
3785 if (block == NULL_TREE) | |
3786 return; | |
3787 | |
3788 block_stack = VEC_alloc (tree, heap, 10); | |
3789 | |
3790 /* Reset the TREE_ASM_WRITTEN bit for all blocks. */ | |
3791 clear_block_marks (block); | |
3792 | |
3793 /* Prune the old trees away, so that they don't get in the way. */ | |
3794 BLOCK_SUBBLOCKS (block) = NULL_TREE; | |
3795 BLOCK_CHAIN (block) = NULL_TREE; | |
3796 | |
3797 /* Recreate the block tree from the note nesting. */ | |
3798 reorder_blocks_1 (get_insns (), block, &block_stack); | |
3799 BLOCK_SUBBLOCKS (block) = blocks_nreverse (BLOCK_SUBBLOCKS (block)); | |
3800 | |
3801 VEC_free (tree, heap, block_stack); | |
3802 } | |
3803 | |
3804 /* Helper function for reorder_blocks. Reset TREE_ASM_WRITTEN. */ | |
3805 | |
3806 void | |
3807 clear_block_marks (tree block) | |
3808 { | |
3809 while (block) | |
3810 { | |
3811 TREE_ASM_WRITTEN (block) = 0; | |
3812 clear_block_marks (BLOCK_SUBBLOCKS (block)); | |
3813 block = BLOCK_CHAIN (block); | |
3814 } | |
3815 } | |
3816 | |
3817 static void | |
3818 reorder_blocks_1 (rtx insns, tree current_block, VEC(tree,heap) **p_block_stack) | |
3819 { | |
3820 rtx insn; | |
3821 | |
3822 for (insn = insns; insn; insn = NEXT_INSN (insn)) | |
3823 { | |
3824 if (NOTE_P (insn)) | |
3825 { | |
3826 if (NOTE_KIND (insn) == NOTE_INSN_BLOCK_BEG) | |
3827 { | |
3828 tree block = NOTE_BLOCK (insn); | |
3829 tree origin; | |
3830 | |
3831 origin = (BLOCK_FRAGMENT_ORIGIN (block) | |
3832 ? BLOCK_FRAGMENT_ORIGIN (block) | |
3833 : block); | |
3834 | |
3835 /* If we have seen this block before, that means it now | |
3836 spans multiple address regions. Create a new fragment. */ | |
3837 if (TREE_ASM_WRITTEN (block)) | |
3838 { | |
3839 tree new_block = copy_node (block); | |
3840 | |
3841 BLOCK_FRAGMENT_ORIGIN (new_block) = origin; | |
3842 BLOCK_FRAGMENT_CHAIN (new_block) | |
3843 = BLOCK_FRAGMENT_CHAIN (origin); | |
3844 BLOCK_FRAGMENT_CHAIN (origin) = new_block; | |
3845 | |
3846 NOTE_BLOCK (insn) = new_block; | |
3847 block = new_block; | |
3848 } | |
3849 | |
3850 BLOCK_SUBBLOCKS (block) = 0; | |
3851 TREE_ASM_WRITTEN (block) = 1; | |
3852 /* When there's only one block for the entire function, | |
3853 current_block == block and we mustn't do this, it | |
3854 will cause infinite recursion. */ | |
3855 if (block != current_block) | |
3856 { | |
3857 if (block != origin) | |
3858 gcc_assert (BLOCK_SUPERCONTEXT (origin) == current_block); | |
3859 | |
3860 BLOCK_SUPERCONTEXT (block) = current_block; | |
3861 BLOCK_CHAIN (block) = BLOCK_SUBBLOCKS (current_block); | |
3862 BLOCK_SUBBLOCKS (current_block) = block; | |
3863 current_block = origin; | |
3864 } | |
3865 VEC_safe_push (tree, heap, *p_block_stack, block); | |
3866 } | |
3867 else if (NOTE_KIND (insn) == NOTE_INSN_BLOCK_END) | |
3868 { | |
3869 NOTE_BLOCK (insn) = VEC_pop (tree, *p_block_stack); | |
3870 BLOCK_SUBBLOCKS (current_block) | |
3871 = blocks_nreverse (BLOCK_SUBBLOCKS (current_block)); | |
3872 current_block = BLOCK_SUPERCONTEXT (current_block); | |
3873 } | |
3874 } | |
3875 } | |
3876 } | |
3877 | |
3878 /* Reverse the order of elements in the chain T of blocks, | |
3879 and return the new head of the chain (old last element). */ | |
3880 | |
3881 tree | |
3882 blocks_nreverse (tree t) | |
3883 { | |
3884 tree prev = 0, decl, next; | |
3885 for (decl = t; decl; decl = next) | |
3886 { | |
3887 next = BLOCK_CHAIN (decl); | |
3888 BLOCK_CHAIN (decl) = prev; | |
3889 prev = decl; | |
3890 } | |
3891 return prev; | |
3892 } | |
3893 | |
3894 /* Count the subblocks of the list starting with BLOCK. If VECTOR is | |
3895 non-NULL, list them all into VECTOR, in a depth-first preorder | |
3896 traversal of the block tree. Also clear TREE_ASM_WRITTEN in all | |
3897 blocks. */ | |
3898 | |
3899 static int | |
3900 all_blocks (tree block, tree *vector) | |
3901 { | |
3902 int n_blocks = 0; | |
3903 | |
3904 while (block) | |
3905 { | |
3906 TREE_ASM_WRITTEN (block) = 0; | |
3907 | |
3908 /* Record this block. */ | |
3909 if (vector) | |
3910 vector[n_blocks] = block; | |
3911 | |
3912 ++n_blocks; | |
3913 | |
3914 /* Record the subblocks, and their subblocks... */ | |
3915 n_blocks += all_blocks (BLOCK_SUBBLOCKS (block), | |
3916 vector ? vector + n_blocks : 0); | |
3917 block = BLOCK_CHAIN (block); | |
3918 } | |
3919 | |
3920 return n_blocks; | |
3921 } | |
3922 | |
3923 /* Return a vector containing all the blocks rooted at BLOCK. The | |
3924 number of elements in the vector is stored in N_BLOCKS_P. The | |
3925 vector is dynamically allocated; it is the caller's responsibility | |
3926 to call `free' on the pointer returned. */ | |
3927 | |
3928 static tree * | |
3929 get_block_vector (tree block, int *n_blocks_p) | |
3930 { | |
3931 tree *block_vector; | |
3932 | |
3933 *n_blocks_p = all_blocks (block, NULL); | |
3934 block_vector = XNEWVEC (tree, *n_blocks_p); | |
3935 all_blocks (block, block_vector); | |
3936 | |
3937 return block_vector; | |
3938 } | |
3939 | |
3940 static GTY(()) int next_block_index = 2; | |
3941 | |
3942 /* Set BLOCK_NUMBER for all the blocks in FN. */ | |
3943 | |
3944 void | |
3945 number_blocks (tree fn) | |
3946 { | |
3947 int i; | |
3948 int n_blocks; | |
3949 tree *block_vector; | |
3950 | |
3951 /* For SDB and XCOFF debugging output, we start numbering the blocks | |
3952 from 1 within each function, rather than keeping a running | |
3953 count. */ | |
3954 #if defined (SDB_DEBUGGING_INFO) || defined (XCOFF_DEBUGGING_INFO) | |
3955 if (write_symbols == SDB_DEBUG || write_symbols == XCOFF_DEBUG) | |
3956 next_block_index = 1; | |
3957 #endif | |
3958 | |
3959 block_vector = get_block_vector (DECL_INITIAL (fn), &n_blocks); | |
3960 | |
3961 /* The top-level BLOCK isn't numbered at all. */ | |
3962 for (i = 1; i < n_blocks; ++i) | |
3963 /* We number the blocks from two. */ | |
3964 BLOCK_NUMBER (block_vector[i]) = next_block_index++; | |
3965 | |
3966 free (block_vector); | |
3967 | |
3968 return; | |
3969 } | |
3970 | |
3971 /* If VAR is present in a subblock of BLOCK, return the subblock. */ | |
3972 | |
3973 tree | |
3974 debug_find_var_in_block_tree (tree var, tree block) | |
3975 { | |
3976 tree t; | |
3977 | |
3978 for (t = BLOCK_VARS (block); t; t = TREE_CHAIN (t)) | |
3979 if (t == var) | |
3980 return block; | |
3981 | |
3982 for (t = BLOCK_SUBBLOCKS (block); t; t = TREE_CHAIN (t)) | |
3983 { | |
3984 tree ret = debug_find_var_in_block_tree (var, t); | |
3985 if (ret) | |
3986 return ret; | |
3987 } | |
3988 | |
3989 return NULL_TREE; | |
3990 } | |
3991 | |
3992 /* Keep track of whether we're in a dummy function context. If we are, | |
3993 we don't want to invoke the set_current_function hook, because we'll | |
3994 get into trouble if the hook calls target_reinit () recursively or | |
3995 when the initial initialization is not yet complete. */ | |
3996 | |
3997 static bool in_dummy_function; | |
3998 | |
3999 /* Invoke the target hook when setting cfun. Update the optimization options | |
4000 if the function uses different options than the default. */ | |
4001 | |
4002 static void | |
4003 invoke_set_current_function_hook (tree fndecl) | |
4004 { | |
4005 if (!in_dummy_function) | |
4006 { | |
4007 tree opts = ((fndecl) | |
4008 ? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl) | |
4009 : optimization_default_node); | |
4010 | |
4011 if (!opts) | |
4012 opts = optimization_default_node; | |
4013 | |
4014 /* Change optimization options if needed. */ | |
4015 if (optimization_current_node != opts) | |
4016 { | |
4017 optimization_current_node = opts; | |
4018 cl_optimization_restore (TREE_OPTIMIZATION (opts)); | |
4019 } | |
4020 | |
4021 targetm.set_current_function (fndecl); | |
4022 } | |
4023 } | |
4024 | |
4025 /* cfun should never be set directly; use this function. */ | |
4026 | |
4027 void | |
4028 set_cfun (struct function *new_cfun) | |
4029 { | |
4030 if (cfun != new_cfun) | |
4031 { | |
4032 cfun = new_cfun; | |
4033 invoke_set_current_function_hook (new_cfun ? new_cfun->decl : NULL_TREE); | |
4034 } | |
4035 } | |
4036 | |
4037 /* Initialized with NOGC, making this poisonous to the garbage collector. */ | |
4038 | |
4039 static VEC(function_p,heap) *cfun_stack; | |
4040 | |
4041 /* Push the current cfun onto the stack, and set cfun to new_cfun. */ | |
4042 | |
4043 void | |
4044 push_cfun (struct function *new_cfun) | |
4045 { | |
4046 VEC_safe_push (function_p, heap, cfun_stack, cfun); | |
4047 set_cfun (new_cfun); | |
4048 } | |
4049 | |
4050 /* Pop cfun from the stack. */ | |
4051 | |
4052 void | |
4053 pop_cfun (void) | |
4054 { | |
4055 struct function *new_cfun = VEC_pop (function_p, cfun_stack); | |
4056 set_cfun (new_cfun); | |
4057 } | |
4058 | |
4059 /* Return value of funcdef and increase it. */ | |
4060 int | |
4061 get_next_funcdef_no (void) | |
4062 { | |
4063 return funcdef_no++; | |
4064 } | |
4065 | |
4066 /* Allocate a function structure for FNDECL and set its contents | |
4067 to the defaults. Set cfun to the newly-allocated object. | |
4068 Some of the helper functions invoked during initialization assume | |
4069 that cfun has already been set. Therefore, assign the new object | |
4070 directly into cfun and invoke the back end hook explicitly at the | |
4071 very end, rather than initializing a temporary and calling set_cfun | |
4072 on it. | |
4073 | |
4074 ABSTRACT_P is true if this is a function that will never be seen by | |
4075 the middle-end. Such functions are front-end concepts (like C++ | |
4076 function templates) that do not correspond directly to functions | |
4077 placed in object files. */ | |
4078 | |
4079 void | |
4080 allocate_struct_function (tree fndecl, bool abstract_p) | |
4081 { | |
4082 tree result; | |
4083 tree fntype = fndecl ? TREE_TYPE (fndecl) : NULL_TREE; | |
4084 | |
4085 cfun = GGC_CNEW (struct function); | |
4086 | |
4087 cfun->function_frequency = FUNCTION_FREQUENCY_NORMAL; | |
4088 | |
4089 init_eh_for_function (); | |
4090 | |
4091 if (init_machine_status) | |
4092 cfun->machine = (*init_machine_status) (); | |
4093 | |
4094 #ifdef OVERRIDE_ABI_FORMAT | |
4095 OVERRIDE_ABI_FORMAT (fndecl); | |
4096 #endif | |
4097 | |
4098 invoke_set_current_function_hook (fndecl); | |
4099 | |
4100 if (fndecl != NULL_TREE) | |
4101 { | |
4102 DECL_STRUCT_FUNCTION (fndecl) = cfun; | |
4103 cfun->decl = fndecl; | |
4104 current_function_funcdef_no = get_next_funcdef_no (); | |
4105 | |
4106 result = DECL_RESULT (fndecl); | |
4107 if (!abstract_p && aggregate_value_p (result, fndecl)) | |
4108 { | |
4109 #ifdef PCC_STATIC_STRUCT_RETURN | |
4110 cfun->returns_pcc_struct = 1; | |
4111 #endif | |
4112 cfun->returns_struct = 1; | |
4113 } | |
4114 | |
4115 cfun->stdarg | |
4116 = (fntype | |
4117 && TYPE_ARG_TYPES (fntype) != 0 | |
4118 && (TREE_VALUE (tree_last (TYPE_ARG_TYPES (fntype))) | |
4119 != void_type_node)); | |
4120 | |
4121 /* Assume all registers in stdarg functions need to be saved. */ | |
4122 cfun->va_list_gpr_size = VA_LIST_MAX_GPR_SIZE; | |
4123 cfun->va_list_fpr_size = VA_LIST_MAX_FPR_SIZE; | |
4124 } | |
4125 } | |
4126 | |
4127 /* This is like allocate_struct_function, but pushes a new cfun for FNDECL | |
4128 instead of just setting it. */ | |
4129 | |
4130 void | |
4131 push_struct_function (tree fndecl) | |
4132 { | |
4133 VEC_safe_push (function_p, heap, cfun_stack, cfun); | |
4134 allocate_struct_function (fndecl, false); | |
4135 } | |
4136 | |
4137 /* Reset cfun, and other non-struct-function variables to defaults as | |
4138 appropriate for emitting rtl at the start of a function. */ | |
4139 | |
4140 static void | |
4141 prepare_function_start (void) | |
4142 { | |
4143 gcc_assert (!crtl->emit.x_last_insn); | |
4144 init_temp_slots (); | |
4145 init_emit (); | |
4146 init_varasm_status (); | |
4147 init_expr (); | |
4148 default_rtl_profile (); | |
4149 | |
4150 cse_not_expected = ! optimize; | |
4151 | |
4152 /* Caller save not needed yet. */ | |
4153 caller_save_needed = 0; | |
4154 | |
4155 /* We haven't done register allocation yet. */ | |
4156 reg_renumber = 0; | |
4157 | |
4158 /* Indicate that we have not instantiated virtual registers yet. */ | |
4159 virtuals_instantiated = 0; | |
4160 | |
4161 /* Indicate that we want CONCATs now. */ | |
4162 generating_concat_p = 1; | |
4163 | |
4164 /* Indicate we have no need of a frame pointer yet. */ | |
4165 frame_pointer_needed = 0; | |
4166 } | |
4167 | |
4168 /* Initialize the rtl expansion mechanism so that we can do simple things | |
4169 like generate sequences. This is used to provide a context during global | |
4170 initialization of some passes. You must call expand_dummy_function_end | |
4171 to exit this context. */ | |
4172 | |
4173 void | |
4174 init_dummy_function_start (void) | |
4175 { | |
4176 gcc_assert (!in_dummy_function); | |
4177 in_dummy_function = true; | |
4178 push_struct_function (NULL_TREE); | |
4179 prepare_function_start (); | |
4180 } | |
4181 | |
4182 /* Generate RTL for the start of the function SUBR (a FUNCTION_DECL tree node) | |
4183 and initialize static variables for generating RTL for the statements | |
4184 of the function. */ | |
4185 | |
4186 void | |
4187 init_function_start (tree subr) | |
4188 { | |
4189 if (subr && DECL_STRUCT_FUNCTION (subr)) | |
4190 set_cfun (DECL_STRUCT_FUNCTION (subr)); | |
4191 else | |
4192 allocate_struct_function (subr, false); | |
4193 prepare_function_start (); | |
4194 | |
4195 /* Warn if this value is an aggregate type, | |
4196 regardless of which calling convention we are using for it. */ | |
4197 if (AGGREGATE_TYPE_P (TREE_TYPE (DECL_RESULT (subr)))) | |
4198 warning (OPT_Waggregate_return, "function returns an aggregate"); | |
4199 } | |
4200 | |
4201 /* Make sure all values used by the optimization passes have sane | |
4202 defaults. */ | |
4203 unsigned int | |
4204 init_function_for_compilation (void) | |
4205 { | |
4206 reg_renumber = 0; | |
4207 | |
4208 /* No prologue/epilogue insns yet. Make sure that these vectors are | |
4209 empty. */ | |
4210 gcc_assert (VEC_length (int, prologue) == 0); | |
4211 gcc_assert (VEC_length (int, epilogue) == 0); | |
4212 gcc_assert (VEC_length (int, sibcall_epilogue) == 0); | |
4213 return 0; | |
4214 } | |
4215 | |
4216 struct rtl_opt_pass pass_init_function = | |
4217 { | |
4218 { | |
4219 RTL_PASS, | |
4220 NULL, /* name */ | |
4221 NULL, /* gate */ | |
4222 init_function_for_compilation, /* execute */ | |
4223 NULL, /* sub */ | |
4224 NULL, /* next */ | |
4225 0, /* static_pass_number */ | |
4226 0, /* tv_id */ | |
4227 0, /* properties_required */ | |
4228 0, /* properties_provided */ | |
4229 0, /* properties_destroyed */ | |
4230 0, /* todo_flags_start */ | |
4231 0 /* todo_flags_finish */ | |
4232 } | |
4233 }; | |
4234 | |
4235 | |
4236 void | |
4237 expand_main_function (void) | |
4238 { | |
4239 #if (defined(INVOKE__main) \ | |
4240 || (!defined(HAS_INIT_SECTION) \ | |
4241 && !defined(INIT_SECTION_ASM_OP) \ | |
4242 && !defined(INIT_ARRAY_SECTION_ASM_OP))) | |
4243 emit_library_call (init_one_libfunc (NAME__MAIN), LCT_NORMAL, VOIDmode, 0); | |
4244 #endif | |
4245 } | |
4246 | |
4247 /* Expand code to initialize the stack_protect_guard. This is invoked at | |
4248 the beginning of a function to be protected. */ | |
4249 | |
4250 #ifndef HAVE_stack_protect_set | |
4251 # define HAVE_stack_protect_set 0 | |
4252 # define gen_stack_protect_set(x,y) (gcc_unreachable (), NULL_RTX) | |
4253 #endif | |
4254 | |
4255 void | |
4256 stack_protect_prologue (void) | |
4257 { | |
4258 tree guard_decl = targetm.stack_protect_guard (); | |
4259 rtx x, y; | |
4260 | |
4261 /* Avoid expand_expr here, because we don't want guard_decl pulled | |
4262 into registers unless absolutely necessary. And we know that | |
4263 crtl->stack_protect_guard is a local stack slot, so this skips | |
4264 all the fluff. */ | |
4265 x = validize_mem (DECL_RTL (crtl->stack_protect_guard)); | |
4266 y = validize_mem (DECL_RTL (guard_decl)); | |
4267 | |
4268 /* Allow the target to copy from Y to X without leaking Y into a | |
4269 register. */ | |
4270 if (HAVE_stack_protect_set) | |
4271 { | |
4272 rtx insn = gen_stack_protect_set (x, y); | |
4273 if (insn) | |
4274 { | |
4275 emit_insn (insn); | |
4276 return; | |
4277 } | |
4278 } | |
4279 | |
4280 /* Otherwise do a straight move. */ | |
4281 emit_move_insn (x, y); | |
4282 } | |
4283 | |
4284 /* Expand code to verify the stack_protect_guard. This is invoked at | |
4285 the end of a function to be protected. */ | |
4286 | |
4287 #ifndef HAVE_stack_protect_test | |
4288 # define HAVE_stack_protect_test 0 | |
4289 # define gen_stack_protect_test(x, y, z) (gcc_unreachable (), NULL_RTX) | |
4290 #endif | |
4291 | |
4292 void | |
4293 stack_protect_epilogue (void) | |
4294 { | |
4295 tree guard_decl = targetm.stack_protect_guard (); | |
4296 rtx label = gen_label_rtx (); | |
4297 rtx x, y, tmp; | |
4298 | |
4299 /* Avoid expand_expr here, because we don't want guard_decl pulled | |
4300 into registers unless absolutely necessary. And we know that | |
4301 crtl->stack_protect_guard is a local stack slot, so this skips | |
4302 all the fluff. */ | |
4303 x = validize_mem (DECL_RTL (crtl->stack_protect_guard)); | |
4304 y = validize_mem (DECL_RTL (guard_decl)); | |
4305 | |
4306 /* Allow the target to compare Y with X without leaking either into | |
4307 a register. */ | |
4308 switch (HAVE_stack_protect_test != 0) | |
4309 { | |
4310 case 1: | |
4311 tmp = gen_stack_protect_test (x, y, label); | |
4312 if (tmp) | |
4313 { | |
4314 emit_insn (tmp); | |
4315 break; | |
4316 } | |
4317 /* FALLTHRU */ | |
4318 | |
4319 default: | |
4320 emit_cmp_and_jump_insns (x, y, EQ, NULL_RTX, ptr_mode, 1, label); | |
4321 break; | |
4322 } | |
4323 | |
4324 /* The noreturn predictor has been moved to the tree level. The rtl-level | |
4325 predictors estimate this branch about 20%, which isn't enough to get | |
4326 things moved out of line. Since this is the only extant case of adding | |
4327 a noreturn function at the rtl level, it doesn't seem worth doing ought | |
4328 except adding the prediction by hand. */ | |
4329 tmp = get_last_insn (); | |
4330 if (JUMP_P (tmp)) | |
4331 predict_insn_def (tmp, PRED_NORETURN, TAKEN); | |
4332 | |
4333 expand_expr_stmt (targetm.stack_protect_fail ()); | |
4334 emit_label (label); | |
4335 } | |
4336 | |
4337 /* Start the RTL for a new function, and set variables used for | |
4338 emitting RTL. | |
4339 SUBR is the FUNCTION_DECL node. | |
4340 PARMS_HAVE_CLEANUPS is nonzero if there are cleanups associated with | |
4341 the function's parameters, which must be run at any return statement. */ | |
4342 | |
4343 void | |
4344 expand_function_start (tree subr) | |
4345 { | |
4346 /* Make sure volatile mem refs aren't considered | |
4347 valid operands of arithmetic insns. */ | |
4348 init_recog_no_volatile (); | |
4349 | |
4350 crtl->profile | |
4351 = (profile_flag | |
4352 && ! DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (subr)); | |
4353 | |
4354 crtl->limit_stack | |
4355 = (stack_limit_rtx != NULL_RTX && ! DECL_NO_LIMIT_STACK (subr)); | |
4356 | |
4357 /* Make the label for return statements to jump to. Do not special | |
4358 case machines with special return instructions -- they will be | |
4359 handled later during jump, ifcvt, or epilogue creation. */ | |
4360 return_label = gen_label_rtx (); | |
4361 | |
4362 /* Initialize rtx used to return the value. */ | |
4363 /* Do this before assign_parms so that we copy the struct value address | |
4364 before any library calls that assign parms might generate. */ | |
4365 | |
4366 /* Decide whether to return the value in memory or in a register. */ | |
4367 if (aggregate_value_p (DECL_RESULT (subr), subr)) | |
4368 { | |
4369 /* Returning something that won't go in a register. */ | |
4370 rtx value_address = 0; | |
4371 | |
4372 #ifdef PCC_STATIC_STRUCT_RETURN | |
4373 if (cfun->returns_pcc_struct) | |
4374 { | |
4375 int size = int_size_in_bytes (TREE_TYPE (DECL_RESULT (subr))); | |
4376 value_address = assemble_static_space (size); | |
4377 } | |
4378 else | |
4379 #endif | |
4380 { | |
4381 rtx sv = targetm.calls.struct_value_rtx (TREE_TYPE (subr), 2); | |
4382 /* Expect to be passed the address of a place to store the value. | |
4383 If it is passed as an argument, assign_parms will take care of | |
4384 it. */ | |
4385 if (sv) | |
4386 { | |
4387 value_address = gen_reg_rtx (Pmode); | |
4388 emit_move_insn (value_address, sv); | |
4389 } | |
4390 } | |
4391 if (value_address) | |
4392 { | |
4393 rtx x = value_address; | |
4394 if (!DECL_BY_REFERENCE (DECL_RESULT (subr))) | |
4395 { | |
4396 x = gen_rtx_MEM (DECL_MODE (DECL_RESULT (subr)), x); | |
4397 set_mem_attributes (x, DECL_RESULT (subr), 1); | |
4398 } | |
4399 SET_DECL_RTL (DECL_RESULT (subr), x); | |
4400 } | |
4401 } | |
4402 else if (DECL_MODE (DECL_RESULT (subr)) == VOIDmode) | |
4403 /* If return mode is void, this decl rtl should not be used. */ | |
4404 SET_DECL_RTL (DECL_RESULT (subr), NULL_RTX); | |
4405 else | |
4406 { | |
4407 /* Compute the return values into a pseudo reg, which we will copy | |
4408 into the true return register after the cleanups are done. */ | |
4409 tree return_type = TREE_TYPE (DECL_RESULT (subr)); | |
4410 if (TYPE_MODE (return_type) != BLKmode | |
4411 && targetm.calls.return_in_msb (return_type)) | |
4412 /* expand_function_end will insert the appropriate padding in | |
4413 this case. Use the return value's natural (unpadded) mode | |
4414 within the function proper. */ | |
4415 SET_DECL_RTL (DECL_RESULT (subr), | |
4416 gen_reg_rtx (TYPE_MODE (return_type))); | |
4417 else | |
4418 { | |
4419 /* In order to figure out what mode to use for the pseudo, we | |
4420 figure out what the mode of the eventual return register will | |
4421 actually be, and use that. */ | |
4422 rtx hard_reg = hard_function_value (return_type, subr, 0, 1); | |
4423 | |
4424 /* Structures that are returned in registers are not | |
4425 aggregate_value_p, so we may see a PARALLEL or a REG. */ | |
4426 if (REG_P (hard_reg)) | |
4427 SET_DECL_RTL (DECL_RESULT (subr), | |
4428 gen_reg_rtx (GET_MODE (hard_reg))); | |
4429 else | |
4430 { | |
4431 gcc_assert (GET_CODE (hard_reg) == PARALLEL); | |
4432 SET_DECL_RTL (DECL_RESULT (subr), gen_group_rtx (hard_reg)); | |
4433 } | |
4434 } | |
4435 | |
4436 /* Set DECL_REGISTER flag so that expand_function_end will copy the | |
4437 result to the real return register(s). */ | |
4438 DECL_REGISTER (DECL_RESULT (subr)) = 1; | |
4439 } | |
4440 | |
4441 /* Initialize rtx for parameters and local variables. | |
4442 In some cases this requires emitting insns. */ | |
4443 assign_parms (subr); | |
4444 | |
4445 /* If function gets a static chain arg, store it. */ | |
4446 if (cfun->static_chain_decl) | |
4447 { | |
4448 tree parm = cfun->static_chain_decl; | |
4449 rtx local = gen_reg_rtx (Pmode); | |
4450 | |
4451 set_decl_incoming_rtl (parm, static_chain_incoming_rtx, false); | |
4452 SET_DECL_RTL (parm, local); | |
4453 mark_reg_pointer (local, TYPE_ALIGN (TREE_TYPE (TREE_TYPE (parm)))); | |
4454 | |
4455 emit_move_insn (local, static_chain_incoming_rtx); | |
4456 } | |
4457 | |
4458 /* If the function receives a non-local goto, then store the | |
4459 bits we need to restore the frame pointer. */ | |
4460 if (cfun->nonlocal_goto_save_area) | |
4461 { | |
4462 tree t_save; | |
4463 rtx r_save; | |
4464 | |
4465 /* ??? We need to do this save early. Unfortunately here is | |
4466 before the frame variable gets declared. Help out... */ | |
4467 tree var = TREE_OPERAND (cfun->nonlocal_goto_save_area, 0); | |
4468 if (!DECL_RTL_SET_P (var)) | |
4469 expand_decl (var); | |
4470 | |
4471 t_save = build4 (ARRAY_REF, ptr_type_node, | |
4472 cfun->nonlocal_goto_save_area, | |
4473 integer_zero_node, NULL_TREE, NULL_TREE); | |
4474 r_save = expand_expr (t_save, NULL_RTX, VOIDmode, EXPAND_WRITE); | |
4475 r_save = convert_memory_address (Pmode, r_save); | |
4476 | |
4477 emit_move_insn (r_save, targetm.builtin_setjmp_frame_value ()); | |
4478 update_nonlocal_goto_save_area (); | |
4479 } | |
4480 | |
4481 /* The following was moved from init_function_start. | |
4482 The move is supposed to make sdb output more accurate. */ | |
4483 /* Indicate the beginning of the function body, | |
4484 as opposed to parm setup. */ | |
4485 emit_note (NOTE_INSN_FUNCTION_BEG); | |
4486 | |
4487 gcc_assert (NOTE_P (get_last_insn ())); | |
4488 | |
4489 parm_birth_insn = get_last_insn (); | |
4490 | |
4491 if (crtl->profile) | |
4492 { | |
4493 #ifdef PROFILE_HOOK | |
4494 PROFILE_HOOK (current_function_funcdef_no); | |
4495 #endif | |
4496 } | |
4497 | |
4498 /* After the display initializations is where the stack checking | |
4499 probe should go. */ | |
4500 if(flag_stack_check) | |
4501 stack_check_probe_note = emit_note (NOTE_INSN_DELETED); | |
4502 | |
4503 /* Make sure there is a line number after the function entry setup code. */ | |
4504 force_next_line_note (); | |
4505 } | |
4506 | |
4507 /* Undo the effects of init_dummy_function_start. */ | |
4508 void | |
4509 expand_dummy_function_end (void) | |
4510 { | |
4511 gcc_assert (in_dummy_function); | |
4512 | |
4513 /* End any sequences that failed to be closed due to syntax errors. */ | |
4514 while (in_sequence_p ()) | |
4515 end_sequence (); | |
4516 | |
4517 /* Outside function body, can't compute type's actual size | |
4518 until next function's body starts. */ | |
4519 | |
4520 free_after_parsing (cfun); | |
4521 free_after_compilation (cfun); | |
4522 pop_cfun (); | |
4523 in_dummy_function = false; | |
4524 } | |
4525 | |
4526 /* Call DOIT for each hard register used as a return value from | |
4527 the current function. */ | |
4528 | |
4529 void | |
4530 diddle_return_value (void (*doit) (rtx, void *), void *arg) | |
4531 { | |
4532 rtx outgoing = crtl->return_rtx; | |
4533 | |
4534 if (! outgoing) | |
4535 return; | |
4536 | |
4537 if (REG_P (outgoing)) | |
4538 (*doit) (outgoing, arg); | |
4539 else if (GET_CODE (outgoing) == PARALLEL) | |
4540 { | |
4541 int i; | |
4542 | |
4543 for (i = 0; i < XVECLEN (outgoing, 0); i++) | |
4544 { | |
4545 rtx x = XEXP (XVECEXP (outgoing, 0, i), 0); | |
4546 | |
4547 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER) | |
4548 (*doit) (x, arg); | |
4549 } | |
4550 } | |
4551 } | |
4552 | |
4553 static void | |
4554 do_clobber_return_reg (rtx reg, void *arg ATTRIBUTE_UNUSED) | |
4555 { | |
4556 emit_clobber (reg); | |
4557 } | |
4558 | |
4559 void | |
4560 clobber_return_register (void) | |
4561 { | |
4562 diddle_return_value (do_clobber_return_reg, NULL); | |
4563 | |
4564 /* In case we do use pseudo to return value, clobber it too. */ | |
4565 if (DECL_RTL_SET_P (DECL_RESULT (current_function_decl))) | |
4566 { | |
4567 tree decl_result = DECL_RESULT (current_function_decl); | |
4568 rtx decl_rtl = DECL_RTL (decl_result); | |
4569 if (REG_P (decl_rtl) && REGNO (decl_rtl) >= FIRST_PSEUDO_REGISTER) | |
4570 { | |
4571 do_clobber_return_reg (decl_rtl, NULL); | |
4572 } | |
4573 } | |
4574 } | |
4575 | |
4576 static void | |
4577 do_use_return_reg (rtx reg, void *arg ATTRIBUTE_UNUSED) | |
4578 { | |
4579 emit_use (reg); | |
4580 } | |
4581 | |
4582 static void | |
4583 use_return_register (void) | |
4584 { | |
4585 diddle_return_value (do_use_return_reg, NULL); | |
4586 } | |
4587 | |
4588 /* Possibly warn about unused parameters. */ | |
4589 void | |
4590 do_warn_unused_parameter (tree fn) | |
4591 { | |
4592 tree decl; | |
4593 | |
4594 for (decl = DECL_ARGUMENTS (fn); | |
4595 decl; decl = TREE_CHAIN (decl)) | |
4596 if (!TREE_USED (decl) && TREE_CODE (decl) == PARM_DECL | |
4597 && DECL_NAME (decl) && !DECL_ARTIFICIAL (decl) | |
4598 && !TREE_NO_WARNING (decl)) | |
4599 warning (OPT_Wunused_parameter, "unused parameter %q+D", decl); | |
4600 } | |
4601 | |
4602 static GTY(()) rtx initial_trampoline; | |
4603 | |
4604 /* Generate RTL for the end of the current function. */ | |
4605 | |
4606 void | |
4607 expand_function_end (void) | |
4608 { | |
4609 rtx clobber_after; | |
4610 | |
4611 /* If arg_pointer_save_area was referenced only from a nested | |
4612 function, we will not have initialized it yet. Do that now. */ | |
4613 if (arg_pointer_save_area && ! crtl->arg_pointer_save_area_init) | |
4614 get_arg_pointer_save_area (); | |
4615 | |
4616 /* If we are doing generic stack checking and this function makes calls, | |
4617 do a stack probe at the start of the function to ensure we have enough | |
4618 space for another stack frame. */ | |
4619 if (flag_stack_check == GENERIC_STACK_CHECK) | |
4620 { | |
4621 rtx insn, seq; | |
4622 | |
4623 for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
4624 if (CALL_P (insn)) | |
4625 { | |
4626 start_sequence (); | |
4627 probe_stack_range (STACK_OLD_CHECK_PROTECT, | |
4628 GEN_INT (STACK_CHECK_MAX_FRAME_SIZE)); | |
4629 seq = get_insns (); | |
4630 end_sequence (); | |
4631 emit_insn_before (seq, stack_check_probe_note); | |
4632 break; | |
4633 } | |
4634 } | |
4635 | |
4636 /* End any sequences that failed to be closed due to syntax errors. */ | |
4637 while (in_sequence_p ()) | |
4638 end_sequence (); | |
4639 | |
4640 clear_pending_stack_adjust (); | |
4641 do_pending_stack_adjust (); | |
4642 | |
4643 /* Output a linenumber for the end of the function. | |
4644 SDB depends on this. */ | |
4645 force_next_line_note (); | |
4646 set_curr_insn_source_location (input_location); | |
4647 | |
4648 /* Before the return label (if any), clobber the return | |
4649 registers so that they are not propagated live to the rest of | |
4650 the function. This can only happen with functions that drop | |
4651 through; if there had been a return statement, there would | |
4652 have either been a return rtx, or a jump to the return label. | |
4653 | |
4654 We delay actual code generation after the current_function_value_rtx | |
4655 is computed. */ | |
4656 clobber_after = get_last_insn (); | |
4657 | |
4658 /* Output the label for the actual return from the function. */ | |
4659 emit_label (return_label); | |
4660 | |
4661 if (USING_SJLJ_EXCEPTIONS) | |
4662 { | |
4663 /* Let except.c know where it should emit the call to unregister | |
4664 the function context for sjlj exceptions. */ | |
4665 if (flag_exceptions) | |
4666 sjlj_emit_function_exit_after (get_last_insn ()); | |
4667 } | |
4668 else | |
4669 { | |
4670 /* We want to ensure that instructions that may trap are not | |
4671 moved into the epilogue by scheduling, because we don't | |
4672 always emit unwind information for the epilogue. */ | |
4673 if (flag_non_call_exceptions) | |
4674 emit_insn (gen_blockage ()); | |
4675 } | |
4676 | |
4677 /* If this is an implementation of throw, do what's necessary to | |
4678 communicate between __builtin_eh_return and the epilogue. */ | |
4679 expand_eh_return (); | |
4680 | |
4681 /* If scalar return value was computed in a pseudo-reg, or was a named | |
4682 return value that got dumped to the stack, copy that to the hard | |
4683 return register. */ | |
4684 if (DECL_RTL_SET_P (DECL_RESULT (current_function_decl))) | |
4685 { | |
4686 tree decl_result = DECL_RESULT (current_function_decl); | |
4687 rtx decl_rtl = DECL_RTL (decl_result); | |
4688 | |
4689 if (REG_P (decl_rtl) | |
4690 ? REGNO (decl_rtl) >= FIRST_PSEUDO_REGISTER | |
4691 : DECL_REGISTER (decl_result)) | |
4692 { | |
4693 rtx real_decl_rtl = crtl->return_rtx; | |
4694 | |
4695 /* This should be set in assign_parms. */ | |
4696 gcc_assert (REG_FUNCTION_VALUE_P (real_decl_rtl)); | |
4697 | |
4698 /* If this is a BLKmode structure being returned in registers, | |
4699 then use the mode computed in expand_return. Note that if | |
4700 decl_rtl is memory, then its mode may have been changed, | |
4701 but that crtl->return_rtx has not. */ | |
4702 if (GET_MODE (real_decl_rtl) == BLKmode) | |
4703 PUT_MODE (real_decl_rtl, GET_MODE (decl_rtl)); | |
4704 | |
4705 /* If a non-BLKmode return value should be padded at the least | |
4706 significant end of the register, shift it left by the appropriate | |
4707 amount. BLKmode results are handled using the group load/store | |
4708 machinery. */ | |
4709 if (TYPE_MODE (TREE_TYPE (decl_result)) != BLKmode | |
4710 && targetm.calls.return_in_msb (TREE_TYPE (decl_result))) | |
4711 { | |
4712 emit_move_insn (gen_rtx_REG (GET_MODE (decl_rtl), | |
4713 REGNO (real_decl_rtl)), | |
4714 decl_rtl); | |
4715 shift_return_value (GET_MODE (decl_rtl), true, real_decl_rtl); | |
4716 } | |
4717 /* If a named return value dumped decl_return to memory, then | |
4718 we may need to re-do the PROMOTE_MODE signed/unsigned | |
4719 extension. */ | |
4720 else if (GET_MODE (real_decl_rtl) != GET_MODE (decl_rtl)) | |
4721 { | |
4722 int unsignedp = TYPE_UNSIGNED (TREE_TYPE (decl_result)); | |
4723 | |
4724 if (targetm.calls.promote_function_return (TREE_TYPE (current_function_decl))) | |
4725 promote_mode (TREE_TYPE (decl_result), GET_MODE (decl_rtl), | |
4726 &unsignedp, 1); | |
4727 | |
4728 convert_move (real_decl_rtl, decl_rtl, unsignedp); | |
4729 } | |
4730 else if (GET_CODE (real_decl_rtl) == PARALLEL) | |
4731 { | |
4732 /* If expand_function_start has created a PARALLEL for decl_rtl, | |
4733 move the result to the real return registers. Otherwise, do | |
4734 a group load from decl_rtl for a named return. */ | |
4735 if (GET_CODE (decl_rtl) == PARALLEL) | |
4736 emit_group_move (real_decl_rtl, decl_rtl); | |
4737 else | |
4738 emit_group_load (real_decl_rtl, decl_rtl, | |
4739 TREE_TYPE (decl_result), | |
4740 int_size_in_bytes (TREE_TYPE (decl_result))); | |
4741 } | |
4742 /* In the case of complex integer modes smaller than a word, we'll | |
4743 need to generate some non-trivial bitfield insertions. Do that | |
4744 on a pseudo and not the hard register. */ | |
4745 else if (GET_CODE (decl_rtl) == CONCAT | |
4746 && GET_MODE_CLASS (GET_MODE (decl_rtl)) == MODE_COMPLEX_INT | |
4747 && GET_MODE_BITSIZE (GET_MODE (decl_rtl)) <= BITS_PER_WORD) | |
4748 { | |
4749 int old_generating_concat_p; | |
4750 rtx tmp; | |
4751 | |
4752 old_generating_concat_p = generating_concat_p; | |
4753 generating_concat_p = 0; | |
4754 tmp = gen_reg_rtx (GET_MODE (decl_rtl)); | |
4755 generating_concat_p = old_generating_concat_p; | |
4756 | |
4757 emit_move_insn (tmp, decl_rtl); | |
4758 emit_move_insn (real_decl_rtl, tmp); | |
4759 } | |
4760 else | |
4761 emit_move_insn (real_decl_rtl, decl_rtl); | |
4762 } | |
4763 } | |
4764 | |
4765 /* If returning a structure, arrange to return the address of the value | |
4766 in a place where debuggers expect to find it. | |
4767 | |
4768 If returning a structure PCC style, | |
4769 the caller also depends on this value. | |
4770 And cfun->returns_pcc_struct is not necessarily set. */ | |
4771 if (cfun->returns_struct | |
4772 || cfun->returns_pcc_struct) | |
4773 { | |
4774 rtx value_address = DECL_RTL (DECL_RESULT (current_function_decl)); | |
4775 tree type = TREE_TYPE (DECL_RESULT (current_function_decl)); | |
4776 rtx outgoing; | |
4777 | |
4778 if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) | |
4779 type = TREE_TYPE (type); | |
4780 else | |
4781 value_address = XEXP (value_address, 0); | |
4782 | |
4783 outgoing = targetm.calls.function_value (build_pointer_type (type), | |
4784 current_function_decl, true); | |
4785 | |
4786 /* Mark this as a function return value so integrate will delete the | |
4787 assignment and USE below when inlining this function. */ | |
4788 REG_FUNCTION_VALUE_P (outgoing) = 1; | |
4789 | |
4790 /* The address may be ptr_mode and OUTGOING may be Pmode. */ | |
4791 value_address = convert_memory_address (GET_MODE (outgoing), | |
4792 value_address); | |
4793 | |
4794 emit_move_insn (outgoing, value_address); | |
4795 | |
4796 /* Show return register used to hold result (in this case the address | |
4797 of the result. */ | |
4798 crtl->return_rtx = outgoing; | |
4799 } | |
4800 | |
4801 /* Emit the actual code to clobber return register. */ | |
4802 { | |
4803 rtx seq; | |
4804 | |
4805 start_sequence (); | |
4806 clobber_return_register (); | |
4807 expand_naked_return (); | |
4808 seq = get_insns (); | |
4809 end_sequence (); | |
4810 | |
4811 emit_insn_after (seq, clobber_after); | |
4812 } | |
4813 | |
4814 /* Output the label for the naked return from the function. */ | |
4815 emit_label (naked_return_label); | |
4816 | |
4817 /* @@@ This is a kludge. We want to ensure that instructions that | |
4818 may trap are not moved into the epilogue by scheduling, because | |
4819 we don't always emit unwind information for the epilogue. */ | |
4820 if (! USING_SJLJ_EXCEPTIONS && flag_non_call_exceptions) | |
4821 emit_insn (gen_blockage ()); | |
4822 | |
4823 /* If stack protection is enabled for this function, check the guard. */ | |
4824 if (crtl->stack_protect_guard) | |
4825 stack_protect_epilogue (); | |
4826 | |
4827 /* If we had calls to alloca, and this machine needs | |
4828 an accurate stack pointer to exit the function, | |
4829 insert some code to save and restore the stack pointer. */ | |
4830 if (! EXIT_IGNORE_STACK | |
4831 && cfun->calls_alloca) | |
4832 { | |
4833 rtx tem = 0; | |
4834 | |
4835 emit_stack_save (SAVE_FUNCTION, &tem, parm_birth_insn); | |
4836 emit_stack_restore (SAVE_FUNCTION, tem, NULL_RTX); | |
4837 } | |
4838 | |
4839 /* ??? This should no longer be necessary since stupid is no longer with | |
4840 us, but there are some parts of the compiler (eg reload_combine, and | |
4841 sh mach_dep_reorg) that still try and compute their own lifetime info | |
4842 instead of using the general framework. */ | |
4843 use_return_register (); | |
4844 } | |
4845 | |
4846 rtx | |
4847 get_arg_pointer_save_area (void) | |
4848 { | |
4849 rtx ret = arg_pointer_save_area; | |
4850 | |
4851 if (! ret) | |
4852 { | |
4853 ret = assign_stack_local (Pmode, GET_MODE_SIZE (Pmode), 0); | |
4854 arg_pointer_save_area = ret; | |
4855 } | |
4856 | |
4857 if (! crtl->arg_pointer_save_area_init) | |
4858 { | |
4859 rtx seq; | |
4860 | |
4861 /* Save the arg pointer at the beginning of the function. The | |
4862 generated stack slot may not be a valid memory address, so we | |
4863 have to check it and fix it if necessary. */ | |
4864 start_sequence (); | |
4865 emit_move_insn (validize_mem (ret), | |
4866 crtl->args.internal_arg_pointer); | |
4867 seq = get_insns (); | |
4868 end_sequence (); | |
4869 | |
4870 push_topmost_sequence (); | |
4871 emit_insn_after (seq, entry_of_function ()); | |
4872 pop_topmost_sequence (); | |
4873 } | |
4874 | |
4875 return ret; | |
4876 } | |
4877 | |
4878 /* Extend a vector that records the INSN_UIDs of INSNS | |
4879 (a list of one or more insns). */ | |
4880 | |
4881 static void | |
4882 record_insns (rtx insns, VEC(int,heap) **vecp) | |
4883 { | |
4884 rtx tmp; | |
4885 | |
4886 for (tmp = insns; tmp != NULL_RTX; tmp = NEXT_INSN (tmp)) | |
4887 VEC_safe_push (int, heap, *vecp, INSN_UID (tmp)); | |
4888 } | |
4889 | |
4890 /* Set the locator of the insn chain starting at INSN to LOC. */ | |
4891 static void | |
4892 set_insn_locators (rtx insn, int loc) | |
4893 { | |
4894 while (insn != NULL_RTX) | |
4895 { | |
4896 if (INSN_P (insn)) | |
4897 INSN_LOCATOR (insn) = loc; | |
4898 insn = NEXT_INSN (insn); | |
4899 } | |
4900 } | |
4901 | |
4902 /* Determine how many INSN_UIDs in VEC are part of INSN. Because we can | |
4903 be running after reorg, SEQUENCE rtl is possible. */ | |
4904 | |
4905 static int | |
4906 contains (const_rtx insn, VEC(int,heap) **vec) | |
4907 { | |
4908 int i, j; | |
4909 | |
4910 if (NONJUMP_INSN_P (insn) | |
4911 && GET_CODE (PATTERN (insn)) == SEQUENCE) | |
4912 { | |
4913 int count = 0; | |
4914 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--) | |
4915 for (j = VEC_length (int, *vec) - 1; j >= 0; --j) | |
4916 if (INSN_UID (XVECEXP (PATTERN (insn), 0, i)) | |
4917 == VEC_index (int, *vec, j)) | |
4918 count++; | |
4919 return count; | |
4920 } | |
4921 else | |
4922 { | |
4923 for (j = VEC_length (int, *vec) - 1; j >= 0; --j) | |
4924 if (INSN_UID (insn) == VEC_index (int, *vec, j)) | |
4925 return 1; | |
4926 } | |
4927 return 0; | |
4928 } | |
4929 | |
4930 int | |
4931 prologue_epilogue_contains (const_rtx insn) | |
4932 { | |
4933 if (contains (insn, &prologue)) | |
4934 return 1; | |
4935 if (contains (insn, &epilogue)) | |
4936 return 1; | |
4937 return 0; | |
4938 } | |
4939 | |
4940 int | |
4941 sibcall_epilogue_contains (const_rtx insn) | |
4942 { | |
4943 if (sibcall_epilogue) | |
4944 return contains (insn, &sibcall_epilogue); | |
4945 return 0; | |
4946 } | |
4947 | |
4948 #ifdef HAVE_return | |
4949 /* Insert gen_return at the end of block BB. This also means updating | |
4950 block_for_insn appropriately. */ | |
4951 | |
4952 static void | |
4953 emit_return_into_block (basic_block bb) | |
4954 { | |
4955 emit_jump_insn_after (gen_return (), BB_END (bb)); | |
4956 } | |
4957 #endif /* HAVE_return */ | |
4958 | |
4959 /* Generate the prologue and epilogue RTL if the machine supports it. Thread | |
4960 this into place with notes indicating where the prologue ends and where | |
4961 the epilogue begins. Update the basic block information when possible. */ | |
4962 | |
4963 static void | |
4964 thread_prologue_and_epilogue_insns (void) | |
4965 { | |
4966 int inserted = 0; | |
4967 edge e; | |
4968 #if defined (HAVE_sibcall_epilogue) || defined (HAVE_epilogue) || defined (HAVE_return) || defined (HAVE_prologue) | |
4969 rtx seq; | |
4970 #endif | |
4971 #if defined (HAVE_epilogue) || defined(HAVE_return) | |
4972 rtx epilogue_end = NULL_RTX; | |
4973 #endif | |
4974 edge_iterator ei; | |
4975 | |
4976 rtl_profile_for_bb (ENTRY_BLOCK_PTR); | |
4977 #ifdef HAVE_prologue | |
4978 if (HAVE_prologue) | |
4979 { | |
4980 start_sequence (); | |
4981 seq = gen_prologue (); | |
4982 emit_insn (seq); | |
4983 | |
4984 /* Insert an explicit USE for the frame pointer | |
4985 if the profiling is on and the frame pointer is required. */ | |
4986 if (crtl->profile && frame_pointer_needed) | |
4987 emit_use (hard_frame_pointer_rtx); | |
4988 | |
4989 /* Retain a map of the prologue insns. */ | |
4990 record_insns (seq, &prologue); | |
4991 emit_note (NOTE_INSN_PROLOGUE_END); | |
4992 | |
4993 #ifndef PROFILE_BEFORE_PROLOGUE | |
4994 /* Ensure that instructions are not moved into the prologue when | |
4995 profiling is on. The call to the profiling routine can be | |
4996 emitted within the live range of a call-clobbered register. */ | |
4997 if (crtl->profile) | |
4998 emit_insn (gen_blockage ()); | |
4999 #endif | |
5000 | |
5001 seq = get_insns (); | |
5002 end_sequence (); | |
5003 set_insn_locators (seq, prologue_locator); | |
5004 | |
5005 /* Can't deal with multiple successors of the entry block | |
5006 at the moment. Function should always have at least one | |
5007 entry point. */ | |
5008 gcc_assert (single_succ_p (ENTRY_BLOCK_PTR)); | |
5009 | |
5010 insert_insn_on_edge (seq, single_succ_edge (ENTRY_BLOCK_PTR)); | |
5011 inserted = 1; | |
5012 } | |
5013 #endif | |
5014 | |
5015 /* If the exit block has no non-fake predecessors, we don't need | |
5016 an epilogue. */ | |
5017 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds) | |
5018 if ((e->flags & EDGE_FAKE) == 0) | |
5019 break; | |
5020 if (e == NULL) | |
5021 goto epilogue_done; | |
5022 | |
5023 rtl_profile_for_bb (EXIT_BLOCK_PTR); | |
5024 #ifdef HAVE_return | |
5025 if (optimize && HAVE_return) | |
5026 { | |
5027 /* If we're allowed to generate a simple return instruction, | |
5028 then by definition we don't need a full epilogue. Examine | |
5029 the block that falls through to EXIT. If it does not | |
5030 contain any code, examine its predecessors and try to | |
5031 emit (conditional) return instructions. */ | |
5032 | |
5033 basic_block last; | |
5034 rtx label; | |
5035 | |
5036 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds) | |
5037 if (e->flags & EDGE_FALLTHRU) | |
5038 break; | |
5039 if (e == NULL) | |
5040 goto epilogue_done; | |
5041 last = e->src; | |
5042 | |
5043 /* Verify that there are no active instructions in the last block. */ | |
5044 label = BB_END (last); | |
5045 while (label && !LABEL_P (label)) | |
5046 { | |
5047 if (active_insn_p (label)) | |
5048 break; | |
5049 label = PREV_INSN (label); | |
5050 } | |
5051 | |
5052 if (BB_HEAD (last) == label && LABEL_P (label)) | |
5053 { | |
5054 edge_iterator ei2; | |
5055 | |
5056 for (ei2 = ei_start (last->preds); (e = ei_safe_edge (ei2)); ) | |
5057 { | |
5058 basic_block bb = e->src; | |
5059 rtx jump; | |
5060 | |
5061 if (bb == ENTRY_BLOCK_PTR) | |
5062 { | |
5063 ei_next (&ei2); | |
5064 continue; | |
5065 } | |
5066 | |
5067 jump = BB_END (bb); | |
5068 if (!JUMP_P (jump) || JUMP_LABEL (jump) != label) | |
5069 { | |
5070 ei_next (&ei2); | |
5071 continue; | |
5072 } | |
5073 | |
5074 /* If we have an unconditional jump, we can replace that | |
5075 with a simple return instruction. */ | |
5076 if (simplejump_p (jump)) | |
5077 { | |
5078 emit_return_into_block (bb); | |
5079 delete_insn (jump); | |
5080 } | |
5081 | |
5082 /* If we have a conditional jump, we can try to replace | |
5083 that with a conditional return instruction. */ | |
5084 else if (condjump_p (jump)) | |
5085 { | |
5086 if (! redirect_jump (jump, 0, 0)) | |
5087 { | |
5088 ei_next (&ei2); | |
5089 continue; | |
5090 } | |
5091 | |
5092 /* If this block has only one successor, it both jumps | |
5093 and falls through to the fallthru block, so we can't | |
5094 delete the edge. */ | |
5095 if (single_succ_p (bb)) | |
5096 { | |
5097 ei_next (&ei2); | |
5098 continue; | |
5099 } | |
5100 } | |
5101 else | |
5102 { | |
5103 ei_next (&ei2); | |
5104 continue; | |
5105 } | |
5106 | |
5107 /* Fix up the CFG for the successful change we just made. */ | |
5108 redirect_edge_succ (e, EXIT_BLOCK_PTR); | |
5109 } | |
5110 | |
5111 /* Emit a return insn for the exit fallthru block. Whether | |
5112 this is still reachable will be determined later. */ | |
5113 | |
5114 emit_barrier_after (BB_END (last)); | |
5115 emit_return_into_block (last); | |
5116 epilogue_end = BB_END (last); | |
5117 single_succ_edge (last)->flags &= ~EDGE_FALLTHRU; | |
5118 goto epilogue_done; | |
5119 } | |
5120 } | |
5121 #endif | |
5122 /* Find the edge that falls through to EXIT. Other edges may exist | |
5123 due to RETURN instructions, but those don't need epilogues. | |
5124 There really shouldn't be a mixture -- either all should have | |
5125 been converted or none, however... */ | |
5126 | |
5127 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds) | |
5128 if (e->flags & EDGE_FALLTHRU) | |
5129 break; | |
5130 if (e == NULL) | |
5131 goto epilogue_done; | |
5132 | |
5133 #ifdef HAVE_epilogue | |
5134 if (HAVE_epilogue) | |
5135 { | |
5136 start_sequence (); | |
5137 epilogue_end = emit_note (NOTE_INSN_EPILOGUE_BEG); | |
5138 seq = gen_epilogue (); | |
5139 emit_jump_insn (seq); | |
5140 | |
5141 /* Retain a map of the epilogue insns. */ | |
5142 record_insns (seq, &epilogue); | |
5143 set_insn_locators (seq, epilogue_locator); | |
5144 | |
5145 seq = get_insns (); | |
5146 end_sequence (); | |
5147 | |
5148 insert_insn_on_edge (seq, e); | |
5149 inserted = 1; | |
5150 } | |
5151 else | |
5152 #endif | |
5153 { | |
5154 basic_block cur_bb; | |
5155 | |
5156 if (! next_active_insn (BB_END (e->src))) | |
5157 goto epilogue_done; | |
5158 /* We have a fall-through edge to the exit block, the source is not | |
5159 at the end of the function, and there will be an assembler epilogue | |
5160 at the end of the function. | |
5161 We can't use force_nonfallthru here, because that would try to | |
5162 use return. Inserting a jump 'by hand' is extremely messy, so | |
5163 we take advantage of cfg_layout_finalize using | |
5164 fixup_fallthru_exit_predecessor. */ | |
5165 cfg_layout_initialize (0); | |
5166 FOR_EACH_BB (cur_bb) | |
5167 if (cur_bb->index >= NUM_FIXED_BLOCKS | |
5168 && cur_bb->next_bb->index >= NUM_FIXED_BLOCKS) | |
5169 cur_bb->aux = cur_bb->next_bb; | |
5170 cfg_layout_finalize (); | |
5171 } | |
5172 epilogue_done: | |
5173 default_rtl_profile (); | |
5174 | |
5175 if (inserted) | |
5176 { | |
5177 commit_edge_insertions (); | |
5178 | |
5179 /* The epilogue insns we inserted may cause the exit edge to no longer | |
5180 be fallthru. */ | |
5181 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds) | |
5182 { | |
5183 if (((e->flags & EDGE_FALLTHRU) != 0) | |
5184 && returnjump_p (BB_END (e->src))) | |
5185 e->flags &= ~EDGE_FALLTHRU; | |
5186 } | |
5187 } | |
5188 | |
5189 #ifdef HAVE_sibcall_epilogue | |
5190 /* Emit sibling epilogues before any sibling call sites. */ | |
5191 for (ei = ei_start (EXIT_BLOCK_PTR->preds); (e = ei_safe_edge (ei)); ) | |
5192 { | |
5193 basic_block bb = e->src; | |
5194 rtx insn = BB_END (bb); | |
5195 | |
5196 if (!CALL_P (insn) | |
5197 || ! SIBLING_CALL_P (insn)) | |
5198 { | |
5199 ei_next (&ei); | |
5200 continue; | |
5201 } | |
5202 | |
5203 start_sequence (); | |
5204 emit_insn (gen_sibcall_epilogue ()); | |
5205 seq = get_insns (); | |
5206 end_sequence (); | |
5207 | |
5208 /* Retain a map of the epilogue insns. Used in life analysis to | |
5209 avoid getting rid of sibcall epilogue insns. Do this before we | |
5210 actually emit the sequence. */ | |
5211 record_insns (seq, &sibcall_epilogue); | |
5212 set_insn_locators (seq, epilogue_locator); | |
5213 | |
5214 emit_insn_before (seq, insn); | |
5215 ei_next (&ei); | |
5216 } | |
5217 #endif | |
5218 | |
5219 #ifdef HAVE_epilogue | |
5220 if (epilogue_end) | |
5221 { | |
5222 rtx insn, next; | |
5223 | |
5224 /* Similarly, move any line notes that appear after the epilogue. | |
5225 There is no need, however, to be quite so anal about the existence | |
5226 of such a note. Also possibly move | |
5227 NOTE_INSN_FUNCTION_BEG notes, as those can be relevant for debug | |
5228 info generation. */ | |
5229 for (insn = epilogue_end; insn; insn = next) | |
5230 { | |
5231 next = NEXT_INSN (insn); | |
5232 if (NOTE_P (insn) | |
5233 && (NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG)) | |
5234 reorder_insns (insn, insn, PREV_INSN (epilogue_end)); | |
5235 } | |
5236 } | |
5237 #endif | |
5238 | |
5239 /* Threading the prologue and epilogue changes the artificial refs | |
5240 in the entry and exit blocks. */ | |
5241 epilogue_completed = 1; | |
5242 df_update_entry_exit_and_calls (); | |
5243 } | |
5244 | |
5245 /* Reposition the prologue-end and epilogue-begin notes after instruction | |
5246 scheduling and delayed branch scheduling. */ | |
5247 | |
5248 void | |
5249 reposition_prologue_and_epilogue_notes (void) | |
5250 { | |
5251 #if defined (HAVE_prologue) || defined (HAVE_epilogue) | |
5252 rtx insn, last, note; | |
5253 int len; | |
5254 | |
5255 if ((len = VEC_length (int, prologue)) > 0) | |
5256 { | |
5257 last = 0, note = 0; | |
5258 | |
5259 /* Scan from the beginning until we reach the last prologue insn. | |
5260 We apparently can't depend on basic_block_{head,end} after | |
5261 reorg has run. */ | |
5262 for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
5263 { | |
5264 if (NOTE_P (insn)) | |
5265 { | |
5266 if (NOTE_KIND (insn) == NOTE_INSN_PROLOGUE_END) | |
5267 note = insn; | |
5268 } | |
5269 else if (contains (insn, &prologue)) | |
5270 { | |
5271 last = insn; | |
5272 if (--len == 0) | |
5273 break; | |
5274 } | |
5275 } | |
5276 | |
5277 if (last) | |
5278 { | |
5279 /* Find the prologue-end note if we haven't already, and | |
5280 move it to just after the last prologue insn. */ | |
5281 if (note == 0) | |
5282 { | |
5283 for (note = last; (note = NEXT_INSN (note));) | |
5284 if (NOTE_P (note) | |
5285 && NOTE_KIND (note) == NOTE_INSN_PROLOGUE_END) | |
5286 break; | |
5287 } | |
5288 | |
5289 /* Avoid placing note between CODE_LABEL and BASIC_BLOCK note. */ | |
5290 if (LABEL_P (last)) | |
5291 last = NEXT_INSN (last); | |
5292 reorder_insns (note, note, last); | |
5293 } | |
5294 } | |
5295 | |
5296 if ((len = VEC_length (int, epilogue)) > 0) | |
5297 { | |
5298 last = 0, note = 0; | |
5299 | |
5300 /* Scan from the end until we reach the first epilogue insn. | |
5301 We apparently can't depend on basic_block_{head,end} after | |
5302 reorg has run. */ | |
5303 for (insn = get_last_insn (); insn; insn = PREV_INSN (insn)) | |
5304 { | |
5305 if (NOTE_P (insn)) | |
5306 { | |
5307 if (NOTE_KIND (insn) == NOTE_INSN_EPILOGUE_BEG) | |
5308 note = insn; | |
5309 } | |
5310 else if (contains (insn, &epilogue)) | |
5311 { | |
5312 last = insn; | |
5313 if (--len == 0) | |
5314 break; | |
5315 } | |
5316 } | |
5317 | |
5318 if (last) | |
5319 { | |
5320 /* Find the epilogue-begin note if we haven't already, and | |
5321 move it to just before the first epilogue insn. */ | |
5322 if (note == 0) | |
5323 { | |
5324 for (note = insn; (note = PREV_INSN (note));) | |
5325 if (NOTE_P (note) | |
5326 && NOTE_KIND (note) == NOTE_INSN_EPILOGUE_BEG) | |
5327 break; | |
5328 } | |
5329 | |
5330 if (PREV_INSN (last) != note) | |
5331 reorder_insns (note, note, PREV_INSN (last)); | |
5332 } | |
5333 } | |
5334 #endif /* HAVE_prologue or HAVE_epilogue */ | |
5335 } | |
5336 | |
5337 /* Returns the name of the current function. */ | |
5338 const char * | |
5339 current_function_name (void) | |
5340 { | |
5341 return lang_hooks.decl_printable_name (cfun->decl, 2); | |
5342 } | |
5343 | |
5344 /* Returns the raw (mangled) name of the current function. */ | |
5345 const char * | |
5346 current_function_assembler_name (void) | |
5347 { | |
5348 return IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (cfun->decl)); | |
5349 } | |
5350 | |
5351 | |
5352 static unsigned int | |
5353 rest_of_handle_check_leaf_regs (void) | |
5354 { | |
5355 #ifdef LEAF_REGISTERS | |
5356 current_function_uses_only_leaf_regs | |
5357 = optimize > 0 && only_leaf_regs_used () && leaf_function_p (); | |
5358 #endif | |
5359 return 0; | |
5360 } | |
5361 | |
5362 /* Insert a TYPE into the used types hash table of CFUN. */ | |
5363 static void | |
5364 used_types_insert_helper (tree type, struct function *func) | |
5365 { | |
5366 if (type != NULL && func != NULL) | |
5367 { | |
5368 void **slot; | |
5369 | |
5370 if (func->used_types_hash == NULL) | |
5371 func->used_types_hash = htab_create_ggc (37, htab_hash_pointer, | |
5372 htab_eq_pointer, NULL); | |
5373 slot = htab_find_slot (func->used_types_hash, type, INSERT); | |
5374 if (*slot == NULL) | |
5375 *slot = type; | |
5376 } | |
5377 } | |
5378 | |
5379 /* Given a type, insert it into the used hash table in cfun. */ | |
5380 void | |
5381 used_types_insert (tree t) | |
5382 { | |
5383 while (POINTER_TYPE_P (t) || TREE_CODE (t) == ARRAY_TYPE) | |
5384 t = TREE_TYPE (t); | |
5385 t = TYPE_MAIN_VARIANT (t); | |
5386 if (debug_info_level > DINFO_LEVEL_NONE) | |
5387 used_types_insert_helper (t, cfun); | |
5388 } | |
5389 | |
5390 struct rtl_opt_pass pass_leaf_regs = | |
5391 { | |
5392 { | |
5393 RTL_PASS, | |
5394 NULL, /* name */ | |
5395 NULL, /* gate */ | |
5396 rest_of_handle_check_leaf_regs, /* execute */ | |
5397 NULL, /* sub */ | |
5398 NULL, /* next */ | |
5399 0, /* static_pass_number */ | |
5400 0, /* tv_id */ | |
5401 0, /* properties_required */ | |
5402 0, /* properties_provided */ | |
5403 0, /* properties_destroyed */ | |
5404 0, /* todo_flags_start */ | |
5405 0 /* todo_flags_finish */ | |
5406 } | |
5407 }; | |
5408 | |
5409 static unsigned int | |
5410 rest_of_handle_thread_prologue_and_epilogue (void) | |
5411 { | |
5412 if (optimize) | |
5413 cleanup_cfg (CLEANUP_EXPENSIVE); | |
5414 /* On some machines, the prologue and epilogue code, or parts thereof, | |
5415 can be represented as RTL. Doing so lets us schedule insns between | |
5416 it and the rest of the code and also allows delayed branch | |
5417 scheduling to operate in the epilogue. */ | |
5418 | |
5419 thread_prologue_and_epilogue_insns (); | |
5420 return 0; | |
5421 } | |
5422 | |
5423 struct rtl_opt_pass pass_thread_prologue_and_epilogue = | |
5424 { | |
5425 { | |
5426 RTL_PASS, | |
5427 "pro_and_epilogue", /* name */ | |
5428 NULL, /* gate */ | |
5429 rest_of_handle_thread_prologue_and_epilogue, /* execute */ | |
5430 NULL, /* sub */ | |
5431 NULL, /* next */ | |
5432 0, /* static_pass_number */ | |
5433 TV_THREAD_PROLOGUE_AND_EPILOGUE, /* tv_id */ | |
5434 0, /* properties_required */ | |
5435 0, /* properties_provided */ | |
5436 0, /* properties_destroyed */ | |
5437 TODO_verify_flow, /* todo_flags_start */ | |
5438 TODO_dump_func | | |
5439 TODO_df_verify | | |
5440 TODO_df_finish | TODO_verify_rtl_sharing | | |
5441 TODO_ggc_collect /* todo_flags_finish */ | |
5442 } | |
5443 }; | |
5444 | |
5445 | |
5446 /* This mini-pass fixes fall-out from SSA in asm statements that have | |
5447 in-out constraints. Say you start with | |
5448 | |
5449 orig = inout; | |
5450 asm ("": "+mr" (inout)); | |
5451 use (orig); | |
5452 | |
5453 which is transformed very early to use explicit output and match operands: | |
5454 | |
5455 orig = inout; | |
5456 asm ("": "=mr" (inout) : "0" (inout)); | |
5457 use (orig); | |
5458 | |
5459 Or, after SSA and copyprop, | |
5460 | |
5461 asm ("": "=mr" (inout_2) : "0" (inout_1)); | |
5462 use (inout_1); | |
5463 | |
5464 Clearly inout_2 and inout_1 can't be coalesced easily anymore, as | |
5465 they represent two separate values, so they will get different pseudo | |
5466 registers during expansion. Then, since the two operands need to match | |
5467 per the constraints, but use different pseudo registers, reload can | |
5468 only register a reload for these operands. But reloads can only be | |
5469 satisfied by hardregs, not by memory, so we need a register for this | |
5470 reload, just because we are presented with non-matching operands. | |
5471 So, even though we allow memory for this operand, no memory can be | |
5472 used for it, just because the two operands don't match. This can | |
5473 cause reload failures on register-starved targets. | |
5474 | |
5475 So it's a symptom of reload not being able to use memory for reloads | |
5476 or, alternatively it's also a symptom of both operands not coming into | |
5477 reload as matching (in which case the pseudo could go to memory just | |
5478 fine, as the alternative allows it, and no reload would be necessary). | |
5479 We fix the latter problem here, by transforming | |
5480 | |
5481 asm ("": "=mr" (inout_2) : "0" (inout_1)); | |
5482 | |
5483 back to | |
5484 | |
5485 inout_2 = inout_1; | |
5486 asm ("": "=mr" (inout_2) : "0" (inout_2)); */ | |
5487 | |
5488 static void | |
5489 match_asm_constraints_1 (rtx insn, rtx *p_sets, int noutputs) | |
5490 { | |
5491 int i; | |
5492 bool changed = false; | |
5493 rtx op = SET_SRC (p_sets[0]); | |
5494 int ninputs = ASM_OPERANDS_INPUT_LENGTH (op); | |
5495 rtvec inputs = ASM_OPERANDS_INPUT_VEC (op); | |
5496 bool *output_matched = XALLOCAVEC (bool, noutputs); | |
5497 | |
5498 memset (output_matched, 0, noutputs * sizeof (bool)); | |
5499 for (i = 0; i < ninputs; i++) | |
5500 { | |
5501 rtx input, output, insns; | |
5502 const char *constraint = ASM_OPERANDS_INPUT_CONSTRAINT (op, i); | |
5503 char *end; | |
5504 int match, j; | |
5505 | |
5506 if (*constraint == '%') | |
5507 constraint++; | |
5508 | |
5509 match = strtoul (constraint, &end, 10); | |
5510 if (end == constraint) | |
5511 continue; | |
5512 | |
5513 gcc_assert (match < noutputs); | |
5514 output = SET_DEST (p_sets[match]); | |
5515 input = RTVEC_ELT (inputs, i); | |
5516 /* Only do the transformation for pseudos. */ | |
5517 if (! REG_P (output) | |
5518 || rtx_equal_p (output, input) | |
5519 || (GET_MODE (input) != VOIDmode | |
5520 && GET_MODE (input) != GET_MODE (output))) | |
5521 continue; | |
5522 | |
5523 /* We can't do anything if the output is also used as input, | |
5524 as we're going to overwrite it. */ | |
5525 for (j = 0; j < ninputs; j++) | |
5526 if (reg_overlap_mentioned_p (output, RTVEC_ELT (inputs, j))) | |
5527 break; | |
5528 if (j != ninputs) | |
5529 continue; | |
5530 | |
5531 /* Avoid changing the same input several times. For | |
5532 asm ("" : "=mr" (out1), "=mr" (out2) : "0" (in), "1" (in)); | |
5533 only change in once (to out1), rather than changing it | |
5534 first to out1 and afterwards to out2. */ | |
5535 if (i > 0) | |
5536 { | |
5537 for (j = 0; j < noutputs; j++) | |
5538 if (output_matched[j] && input == SET_DEST (p_sets[j])) | |
5539 break; | |
5540 if (j != noutputs) | |
5541 continue; | |
5542 } | |
5543 output_matched[match] = true; | |
5544 | |
5545 start_sequence (); | |
5546 emit_move_insn (output, input); | |
5547 insns = get_insns (); | |
5548 end_sequence (); | |
5549 emit_insn_before (insns, insn); | |
5550 | |
5551 /* Now replace all mentions of the input with output. We can't | |
5552 just replace the occurrence in inputs[i], as the register might | |
5553 also be used in some other input (or even in an address of an | |
5554 output), which would mean possibly increasing the number of | |
5555 inputs by one (namely 'output' in addition), which might pose | |
5556 a too complicated problem for reload to solve. E.g. this situation: | |
5557 | |
5558 asm ("" : "=r" (output), "=m" (input) : "0" (input)) | |
5559 | |
5560 Here 'input' is used in two occurrences as input (once for the | |
5561 input operand, once for the address in the second output operand). | |
5562 If we would replace only the occurrence of the input operand (to | |
5563 make the matching) we would be left with this: | |
5564 | |
5565 output = input | |
5566 asm ("" : "=r" (output), "=m" (input) : "0" (output)) | |
5567 | |
5568 Now we suddenly have two different input values (containing the same | |
5569 value, but different pseudos) where we formerly had only one. | |
5570 With more complicated asms this might lead to reload failures | |
5571 which wouldn't have happen without this pass. So, iterate over | |
5572 all operands and replace all occurrences of the register used. */ | |
5573 for (j = 0; j < noutputs; j++) | |
5574 if (!rtx_equal_p (SET_DEST (p_sets[j]), input) | |
5575 && reg_overlap_mentioned_p (input, SET_DEST (p_sets[j]))) | |
5576 SET_DEST (p_sets[j]) = replace_rtx (SET_DEST (p_sets[j]), | |
5577 input, output); | |
5578 for (j = 0; j < ninputs; j++) | |
5579 if (reg_overlap_mentioned_p (input, RTVEC_ELT (inputs, j))) | |
5580 RTVEC_ELT (inputs, j) = replace_rtx (RTVEC_ELT (inputs, j), | |
5581 input, output); | |
5582 | |
5583 changed = true; | |
5584 } | |
5585 | |
5586 if (changed) | |
5587 df_insn_rescan (insn); | |
5588 } | |
5589 | |
5590 static unsigned | |
5591 rest_of_match_asm_constraints (void) | |
5592 { | |
5593 basic_block bb; | |
5594 rtx insn, pat, *p_sets; | |
5595 int noutputs; | |
5596 | |
5597 if (!crtl->has_asm_statement) | |
5598 return 0; | |
5599 | |
5600 df_set_flags (DF_DEFER_INSN_RESCAN); | |
5601 FOR_EACH_BB (bb) | |
5602 { | |
5603 FOR_BB_INSNS (bb, insn) | |
5604 { | |
5605 if (!INSN_P (insn)) | |
5606 continue; | |
5607 | |
5608 pat = PATTERN (insn); | |
5609 if (GET_CODE (pat) == PARALLEL) | |
5610 p_sets = &XVECEXP (pat, 0, 0), noutputs = XVECLEN (pat, 0); | |
5611 else if (GET_CODE (pat) == SET) | |
5612 p_sets = &PATTERN (insn), noutputs = 1; | |
5613 else | |
5614 continue; | |
5615 | |
5616 if (GET_CODE (*p_sets) == SET | |
5617 && GET_CODE (SET_SRC (*p_sets)) == ASM_OPERANDS) | |
5618 match_asm_constraints_1 (insn, p_sets, noutputs); | |
5619 } | |
5620 } | |
5621 | |
5622 return TODO_df_finish; | |
5623 } | |
5624 | |
5625 struct rtl_opt_pass pass_match_asm_constraints = | |
5626 { | |
5627 { | |
5628 RTL_PASS, | |
5629 "asmcons", /* name */ | |
5630 NULL, /* gate */ | |
5631 rest_of_match_asm_constraints, /* execute */ | |
5632 NULL, /* sub */ | |
5633 NULL, /* next */ | |
5634 0, /* static_pass_number */ | |
5635 0, /* tv_id */ | |
5636 0, /* properties_required */ | |
5637 0, /* properties_provided */ | |
5638 0, /* properties_destroyed */ | |
5639 0, /* todo_flags_start */ | |
5640 TODO_dump_func /* todo_flags_finish */ | |
5641 } | |
5642 }; | |
5643 | |
5644 | |
5645 #include "gt-function.h" |