1 | /* Expands front end tree to back end RTL for GCC. |
2 | Copyright (C) 1987-2023 Free Software Foundation, Inc. |
3 | |
4 | This file is part of GCC. |
5 | |
6 | GCC is free software; you can redistribute it and/or modify it under |
7 | the terms of the GNU General Public License as published by the Free |
8 | Software Foundation; either version 3, or (at your option) any later |
9 | version. |
10 | |
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
14 | for more details. |
15 | |
16 | You should have received a copy of the GNU General Public License |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | /* This file handles the generation of rtl code from tree structure |
21 | at the level of the function as a whole. |
22 | It creates the rtl expressions for parameters and auto variables |
23 | and has full responsibility for allocating stack slots. |
24 | |
25 | `expand_function_start' is called at the beginning of a function, |
26 | before the function body is parsed, and `expand_function_end' is |
27 | called after parsing the body. |
28 | |
29 | Call `assign_stack_local' to allocate a stack slot for a local variable. |
30 | This is usually done during the RTL generation for the function body, |
31 | but it can also be done in the reload pass when a pseudo-register does |
32 | not get a hard register. */ |
33 | |
34 | #include "config.h" |
35 | #include "system.h" |
36 | #include "coretypes.h" |
37 | #include "backend.h" |
38 | #include "target.h" |
39 | #include "rtl.h" |
40 | #include "tree.h" |
41 | #include "gimple-expr.h" |
42 | #include "cfghooks.h" |
43 | #include "df.h" |
44 | #include "memmodel.h" |
45 | #include "tm_p.h" |
46 | #include "stringpool.h" |
47 | #include "expmed.h" |
48 | #include "optabs.h" |
49 | #include "opts.h" |
50 | #include "regs.h" |
51 | #include "emit-rtl.h" |
52 | #include "recog.h" |
53 | #include "rtl-error.h" |
54 | #include "hard-reg-set.h" |
55 | #include "alias.h" |
56 | #include "fold-const.h" |
57 | #include "stor-layout.h" |
58 | #include "varasm.h" |
59 | #include "except.h" |
60 | #include "dojump.h" |
61 | #include "explow.h" |
62 | #include "calls.h" |
63 | #include "expr.h" |
64 | #include "optabs-tree.h" |
65 | #include "output.h" |
66 | #include "langhooks.h" |
67 | #include "common/common-target.h" |
68 | #include "gimplify.h" |
69 | #include "tree-pass.h" |
70 | #include "cfgrtl.h" |
71 | #include "cfganal.h" |
72 | #include "cfgbuild.h" |
73 | #include "cfgcleanup.h" |
74 | #include "cfgexpand.h" |
75 | #include "shrink-wrap.h" |
76 | #include "toplev.h" |
77 | #include "rtl-iter.h" |
78 | #include "tree-dfa.h" |
79 | #include "tree-ssa.h" |
80 | #include "stringpool.h" |
81 | #include "attribs.h" |
82 | #include "gimple.h" |
83 | #include "options.h" |
84 | #include "function-abi.h" |
85 | #include "value-range.h" |
86 | #include "gimple-range.h" |
87 | |
88 | /* So we can assign to cfun in this file. */ |
89 | #undef cfun |
90 | |
91 | #ifndef STACK_ALIGNMENT_NEEDED |
92 | #define STACK_ALIGNMENT_NEEDED 1 |
93 | #endif |
94 | |
95 | #define STACK_BYTES (STACK_BOUNDARY / BITS_PER_UNIT) |
96 | |
97 | /* Round a value to the lowest integer less than it that is a multiple of |
98 | the required alignment. Avoid using division in case the value is |
99 | negative. Assume the alignment is a power of two. */ |
100 | #define FLOOR_ROUND(VALUE,ALIGN) ((VALUE) & ~((ALIGN) - 1)) |
101 | |
102 | /* Similar, but round to the next highest integer that meets the |
103 | alignment. */ |
104 | #define CEIL_ROUND(VALUE,ALIGN) (((VALUE) + (ALIGN) - 1) & ~((ALIGN)- 1)) |
105 | |
106 | /* Nonzero once virtual register instantiation has been done. |
107 | assign_stack_local uses frame_pointer_rtx when this is nonzero. |
108 | calls.cc:emit_library_call_value_1 uses it to set up |
109 | post-instantiation libcalls. */ |
110 | int virtuals_instantiated; |
111 | |
112 | /* Assign unique numbers to labels generated for profiling, debugging, etc. */ |
113 | static GTY(()) int funcdef_no; |
114 | |
115 | /* These variables hold pointers to functions to create and destroy |
116 | target specific, per-function data structures. */ |
117 | struct machine_function * (*init_machine_status) (void); |
118 | |
119 | /* The currently compiled function. */ |
120 | struct function *cfun = 0; |
121 | |
122 | /* These hashes record the prologue and epilogue insns. */ |
123 | |
124 | struct insn_cache_hasher : ggc_cache_ptr_hash<rtx_def> |
125 | { |
126 | static hashval_t hash (rtx x) { return htab_hash_pointer (x); } |
127 | static bool equal (rtx a, rtx b) { return a == b; } |
128 | }; |
129 | |
130 | static GTY((cache)) |
131 | hash_table<insn_cache_hasher> *prologue_insn_hash; |
132 | static GTY((cache)) |
133 | hash_table<insn_cache_hasher> *epilogue_insn_hash; |
134 | |
135 | |
136 | hash_table<used_type_hasher> *types_used_by_vars_hash = NULL; |
137 | vec<tree, va_gc> *types_used_by_cur_var_decl; |
138 | |
139 | /* Forward declarations. */ |
140 | |
141 | static class temp_slot *find_temp_slot_from_address (rtx); |
142 | static void pad_to_arg_alignment (struct args_size *, int, struct args_size *); |
143 | static void pad_below (struct args_size *, machine_mode, tree); |
144 | static void reorder_blocks_1 (rtx_insn *, tree, vec<tree> *); |
145 | static int all_blocks (tree, tree *); |
146 | static tree *get_block_vector (tree, int *); |
147 | extern tree debug_find_var_in_block_tree (tree, tree); |
148 | /* We always define `record_insns' even if it's not used so that we |
149 | can always export `prologue_epilogue_contains'. */ |
150 | static void record_insns (rtx_insn *, rtx, hash_table<insn_cache_hasher> **) |
151 | ATTRIBUTE_UNUSED; |
152 | static bool contains (const rtx_insn *, hash_table<insn_cache_hasher> *); |
153 | static void prepare_function_start (void); |
154 | static void do_clobber_return_reg (rtx, void *); |
155 | static void do_use_return_reg (rtx, void *); |
156 | |
157 | |
158 | /* Stack of nested functions. */ |
159 | /* Keep track of the cfun stack. */ |
160 | |
161 | static vec<function *> function_context_stack; |
162 | |
163 | /* Save the current context for compilation of a nested function. |
164 | This is called from language-specific code. */ |
165 | |
166 | void |
167 | push_function_context (void) |
168 | { |
169 | if (cfun == 0) |
170 | allocate_struct_function (NULL, false); |
171 | |
172 | function_context_stack.safe_push (obj: cfun); |
173 | set_cfun (NULL); |
174 | } |
175 | |
176 | /* Restore the last saved context, at the end of a nested function. |
177 | This function is called from language-specific code. */ |
178 | |
179 | void |
180 | pop_function_context (void) |
181 | { |
182 | struct function *p = function_context_stack.pop (); |
183 | set_cfun (new_cfun: p); |
184 | current_function_decl = p->decl; |
185 | |
186 | /* Reset variables that have known state during rtx generation. */ |
187 | virtuals_instantiated = 0; |
188 | generating_concat_p = 1; |
189 | } |
190 | |
191 | /* Clear out all parts of the state in F that can safely be discarded |
192 | after the function has been parsed, but not compiled, to let |
193 | garbage collection reclaim the memory. */ |
194 | |
195 | void |
196 | free_after_parsing (struct function *f) |
197 | { |
198 | f->language = 0; |
199 | } |
200 | |
201 | /* Clear out all parts of the state in F that can safely be discarded |
202 | after the function has been compiled, to let garbage collection |
203 | reclaim the memory. */ |
204 | |
205 | void |
206 | free_after_compilation (struct function *f) |
207 | { |
208 | prologue_insn_hash = NULL; |
209 | epilogue_insn_hash = NULL; |
210 | |
211 | free (crtl->emit.regno_pointer_align); |
212 | |
213 | memset (crtl, c: 0, n: sizeof (struct rtl_data)); |
214 | f->eh = NULL; |
215 | f->machine = NULL; |
216 | f->cfg = NULL; |
217 | f->curr_properties &= ~PROP_cfg; |
218 | |
219 | regno_reg_rtx = NULL; |
220 | } |
221 | |
222 | /* Return size needed for stack frame based on slots so far allocated. |
223 | This size counts from zero. It is not rounded to PREFERRED_STACK_BOUNDARY; |
224 | the caller may have to do that. */ |
225 | |
226 | poly_int64 |
227 | get_frame_size (void) |
228 | { |
229 | if (FRAME_GROWS_DOWNWARD) |
230 | return -frame_offset; |
231 | else |
232 | return frame_offset; |
233 | } |
234 | |
235 | /* Issue an error message and return TRUE if frame OFFSET overflows in |
236 | the signed target pointer arithmetics for function FUNC. Otherwise |
237 | return FALSE. */ |
238 | |
239 | bool |
240 | frame_offset_overflow (poly_int64 offset, tree func) |
241 | { |
242 | poly_uint64 size = FRAME_GROWS_DOWNWARD ? -offset : offset; |
243 | unsigned HOST_WIDE_INT limit |
244 | = ((HOST_WIDE_INT_1U << (GET_MODE_BITSIZE (Pmode) - 1)) |
245 | /* Leave room for the fixed part of the frame. */ |
246 | - 64 * UNITS_PER_WORD); |
247 | |
248 | if (!coeffs_in_range_p (a: size, b: 0U, c: limit)) |
249 | { |
250 | unsigned HOST_WIDE_INT hwisize; |
251 | if (size.is_constant (const_value: &hwisize)) |
252 | error_at (DECL_SOURCE_LOCATION (func), |
253 | "total size of local objects %wu exceeds maximum %wu" , |
254 | hwisize, limit); |
255 | else |
256 | error_at (DECL_SOURCE_LOCATION (func), |
257 | "total size of local objects exceeds maximum %wu" , |
258 | limit); |
259 | return true; |
260 | } |
261 | |
262 | return false; |
263 | } |
264 | |
265 | /* Return the minimum spill slot alignment for a register of mode MODE. */ |
266 | |
267 | unsigned int |
268 | spill_slot_alignment (machine_mode mode ATTRIBUTE_UNUSED) |
269 | { |
270 | return STACK_SLOT_ALIGNMENT (NULL_TREE, mode, GET_MODE_ALIGNMENT (mode)); |
271 | } |
272 | |
273 | /* Return stack slot alignment in bits for TYPE and MODE. */ |
274 | |
275 | static unsigned int |
276 | get_stack_local_alignment (tree type, machine_mode mode) |
277 | { |
278 | unsigned int alignment; |
279 | |
280 | if (mode == BLKmode) |
281 | alignment = BIGGEST_ALIGNMENT; |
282 | else |
283 | alignment = GET_MODE_ALIGNMENT (mode); |
284 | |
285 | /* Allow the frond-end to (possibly) increase the alignment of this |
286 | stack slot. */ |
287 | if (! type) |
288 | type = lang_hooks.types.type_for_mode (mode, 0); |
289 | |
290 | return STACK_SLOT_ALIGNMENT (type, mode, alignment); |
291 | } |
292 | |
293 | /* Determine whether it is possible to fit a stack slot of size SIZE and |
294 | alignment ALIGNMENT into an area in the stack frame that starts at |
295 | frame offset START and has a length of LENGTH. If so, store the frame |
296 | offset to be used for the stack slot in *POFFSET and return true; |
297 | return false otherwise. This function will extend the frame size when |
298 | given a start/length pair that lies at the end of the frame. */ |
299 | |
300 | static bool |
301 | try_fit_stack_local (poly_int64 start, poly_int64 length, |
302 | poly_int64 size, unsigned int alignment, |
303 | poly_int64 *poffset) |
304 | { |
305 | poly_int64 this_frame_offset; |
306 | int frame_off, frame_alignment, frame_phase; |
307 | |
308 | /* Calculate how many bytes the start of local variables is off from |
309 | stack alignment. */ |
310 | frame_alignment = PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT; |
311 | frame_off = targetm.starting_frame_offset () % frame_alignment; |
312 | frame_phase = frame_off ? frame_alignment - frame_off : 0; |
313 | |
314 | /* Round the frame offset to the specified alignment. */ |
315 | |
316 | if (FRAME_GROWS_DOWNWARD) |
317 | this_frame_offset |
318 | = (aligned_lower_bound (value: start + length - size - frame_phase, align: alignment) |
319 | + frame_phase); |
320 | else |
321 | this_frame_offset |
322 | = aligned_upper_bound (value: start - frame_phase, align: alignment) + frame_phase; |
323 | |
324 | /* See if it fits. If this space is at the edge of the frame, |
325 | consider extending the frame to make it fit. Our caller relies on |
326 | this when allocating a new slot. */ |
327 | if (maybe_lt (a: this_frame_offset, b: start)) |
328 | { |
329 | if (known_eq (frame_offset, start)) |
330 | frame_offset = this_frame_offset; |
331 | else |
332 | return false; |
333 | } |
334 | else if (maybe_gt (this_frame_offset + size, start + length)) |
335 | { |
336 | if (known_eq (frame_offset, start + length)) |
337 | frame_offset = this_frame_offset + size; |
338 | else |
339 | return false; |
340 | } |
341 | |
342 | *poffset = this_frame_offset; |
343 | return true; |
344 | } |
345 | |
346 | /* Create a new frame_space structure describing free space in the stack |
347 | frame beginning at START and ending at END, and chain it into the |
348 | function's frame_space_list. */ |
349 | |
350 | static void |
351 | add_frame_space (poly_int64 start, poly_int64 end) |
352 | { |
353 | class frame_space *space = ggc_alloc<frame_space> (); |
354 | space->next = crtl->frame_space_list; |
355 | crtl->frame_space_list = space; |
356 | space->start = start; |
357 | space->length = end - start; |
358 | } |
359 | |
360 | /* Allocate a stack slot of SIZE bytes and return a MEM rtx for it |
361 | with machine mode MODE. |
362 | |
363 | ALIGN controls the amount of alignment for the address of the slot: |
364 | 0 means according to MODE, |
365 | -1 means use BIGGEST_ALIGNMENT and round size to multiple of that, |
366 | -2 means use BITS_PER_UNIT, |
367 | positive specifies alignment boundary in bits. |
368 | |
369 | KIND has ASLK_REDUCE_ALIGN bit set if it is OK to reduce |
370 | alignment and ASLK_RECORD_PAD bit set if we should remember |
371 | extra space we allocated for alignment purposes. When we are |
372 | called from assign_stack_temp_for_type, it is not set so we don't |
373 | track the same stack slot in two independent lists. |
374 | |
375 | We do not round to stack_boundary here. */ |
376 | |
377 | rtx |
378 | assign_stack_local_1 (machine_mode mode, poly_int64 size, |
379 | int align, int kind) |
380 | { |
381 | rtx x, addr; |
382 | poly_int64 bigend_correction = 0; |
383 | poly_int64 slot_offset = 0, old_frame_offset; |
384 | unsigned int alignment, alignment_in_bits; |
385 | |
386 | if (align == 0) |
387 | { |
388 | alignment = get_stack_local_alignment (NULL, mode); |
389 | alignment /= BITS_PER_UNIT; |
390 | } |
391 | else if (align == -1) |
392 | { |
393 | alignment = BIGGEST_ALIGNMENT / BITS_PER_UNIT; |
394 | size = aligned_upper_bound (value: size, align: alignment); |
395 | } |
396 | else if (align == -2) |
397 | alignment = 1; /* BITS_PER_UNIT / BITS_PER_UNIT */ |
398 | else |
399 | alignment = align / BITS_PER_UNIT; |
400 | |
401 | alignment_in_bits = alignment * BITS_PER_UNIT; |
402 | |
403 | /* Ignore alignment if it exceeds MAX_SUPPORTED_STACK_ALIGNMENT. */ |
404 | if (alignment_in_bits > MAX_SUPPORTED_STACK_ALIGNMENT) |
405 | { |
406 | alignment_in_bits = MAX_SUPPORTED_STACK_ALIGNMENT; |
407 | alignment = MAX_SUPPORTED_STACK_ALIGNMENT / BITS_PER_UNIT; |
408 | } |
409 | |
410 | if (SUPPORTS_STACK_ALIGNMENT) |
411 | { |
412 | if (crtl->stack_alignment_estimated < alignment_in_bits) |
413 | { |
414 | if (!crtl->stack_realign_processed) |
415 | crtl->stack_alignment_estimated = alignment_in_bits; |
416 | else |
417 | { |
418 | /* If stack is realigned and stack alignment value |
419 | hasn't been finalized, it is OK not to increase |
420 | stack_alignment_estimated. The bigger alignment |
421 | requirement is recorded in stack_alignment_needed |
422 | below. */ |
423 | gcc_assert (!crtl->stack_realign_finalized); |
424 | if (!crtl->stack_realign_needed) |
425 | { |
426 | /* It is OK to reduce the alignment as long as the |
427 | requested size is 0 or the estimated stack |
428 | alignment >= mode alignment. */ |
429 | gcc_assert ((kind & ASLK_REDUCE_ALIGN) |
430 | || known_eq (size, 0) |
431 | || (crtl->stack_alignment_estimated |
432 | >= GET_MODE_ALIGNMENT (mode))); |
433 | alignment_in_bits = crtl->stack_alignment_estimated; |
434 | alignment = alignment_in_bits / BITS_PER_UNIT; |
435 | } |
436 | } |
437 | } |
438 | } |
439 | |
440 | if (crtl->stack_alignment_needed < alignment_in_bits) |
441 | crtl->stack_alignment_needed = alignment_in_bits; |
442 | if (crtl->max_used_stack_slot_alignment < alignment_in_bits) |
443 | crtl->max_used_stack_slot_alignment = alignment_in_bits; |
444 | |
445 | if (mode != BLKmode || maybe_ne (a: size, b: 0)) |
446 | { |
447 | if (kind & ASLK_RECORD_PAD) |
448 | { |
449 | class frame_space **psp; |
450 | |
451 | for (psp = &crtl->frame_space_list; *psp; psp = &(*psp)->next) |
452 | { |
453 | class frame_space *space = *psp; |
454 | if (!try_fit_stack_local (start: space->start, length: space->length, size, |
455 | alignment, poffset: &slot_offset)) |
456 | continue; |
457 | *psp = space->next; |
458 | if (known_gt (slot_offset, space->start)) |
459 | add_frame_space (start: space->start, end: slot_offset); |
460 | if (known_lt (slot_offset + size, space->start + space->length)) |
461 | add_frame_space (start: slot_offset + size, |
462 | end: space->start + space->length); |
463 | goto found_space; |
464 | } |
465 | } |
466 | } |
467 | else if (!STACK_ALIGNMENT_NEEDED) |
468 | { |
469 | slot_offset = frame_offset; |
470 | goto found_space; |
471 | } |
472 | |
473 | old_frame_offset = frame_offset; |
474 | |
475 | if (FRAME_GROWS_DOWNWARD) |
476 | { |
477 | frame_offset -= size; |
478 | try_fit_stack_local (frame_offset, length: size, size, alignment, poffset: &slot_offset); |
479 | |
480 | if (kind & ASLK_RECORD_PAD) |
481 | { |
482 | if (known_gt (slot_offset, frame_offset)) |
483 | add_frame_space (frame_offset, end: slot_offset); |
484 | if (known_lt (slot_offset + size, old_frame_offset)) |
485 | add_frame_space (start: slot_offset + size, end: old_frame_offset); |
486 | } |
487 | } |
488 | else |
489 | { |
490 | frame_offset += size; |
491 | try_fit_stack_local (start: old_frame_offset, length: size, size, alignment, poffset: &slot_offset); |
492 | |
493 | if (kind & ASLK_RECORD_PAD) |
494 | { |
495 | if (known_gt (slot_offset, old_frame_offset)) |
496 | add_frame_space (start: old_frame_offset, end: slot_offset); |
497 | if (known_lt (slot_offset + size, frame_offset)) |
498 | add_frame_space (start: slot_offset + size, frame_offset); |
499 | } |
500 | } |
501 | |
502 | found_space: |
503 | /* On a big-endian machine, if we are allocating more space than we will use, |
504 | use the least significant bytes of those that are allocated. */ |
505 | if (mode != BLKmode) |
506 | { |
507 | /* The slot size can sometimes be smaller than the mode size; |
508 | e.g. the rs6000 port allocates slots with a vector mode |
509 | that have the size of only one element. However, the slot |
510 | size must always be ordered wrt to the mode size, in the |
511 | same way as for a subreg. */ |
512 | gcc_checking_assert (ordered_p (GET_MODE_SIZE (mode), size)); |
513 | if (BYTES_BIG_ENDIAN && maybe_lt (a: GET_MODE_SIZE (mode), b: size)) |
514 | bigend_correction = size - GET_MODE_SIZE (mode); |
515 | } |
516 | |
517 | /* If we have already instantiated virtual registers, return the actual |
518 | address relative to the frame pointer. */ |
519 | if (virtuals_instantiated) |
520 | addr = plus_constant (Pmode, frame_pointer_rtx, |
521 | trunc_int_for_mode |
522 | (slot_offset + bigend_correction |
523 | + targetm.starting_frame_offset (), Pmode)); |
524 | else |
525 | addr = plus_constant (Pmode, virtual_stack_vars_rtx, |
526 | trunc_int_for_mode |
527 | (slot_offset + bigend_correction, |
528 | Pmode)); |
529 | |
530 | x = gen_rtx_MEM (mode, addr); |
531 | set_mem_align (x, alignment_in_bits); |
532 | MEM_NOTRAP_P (x) = 1; |
533 | |
534 | vec_safe_push (stack_slot_list, obj: x); |
535 | |
536 | if (frame_offset_overflow (frame_offset, func: current_function_decl)) |
537 | frame_offset = 0; |
538 | |
539 | return x; |
540 | } |
541 | |
542 | /* Wrap up assign_stack_local_1 with last parameter as false. */ |
543 | |
544 | rtx |
545 | assign_stack_local (machine_mode mode, poly_int64 size, int align) |
546 | { |
547 | return assign_stack_local_1 (mode, size, align, ASLK_RECORD_PAD); |
548 | } |
549 | |
550 | /* In order to evaluate some expressions, such as function calls returning |
551 | structures in memory, we need to temporarily allocate stack locations. |
552 | We record each allocated temporary in the following structure. |
553 | |
554 | Associated with each temporary slot is a nesting level. When we pop up |
555 | one level, all temporaries associated with the previous level are freed. |
556 | Normally, all temporaries are freed after the execution of the statement |
557 | in which they were created. However, if we are inside a ({...}) grouping, |
558 | the result may be in a temporary and hence must be preserved. If the |
559 | result could be in a temporary, we preserve it if we can determine which |
560 | one it is in. If we cannot determine which temporary may contain the |
561 | result, all temporaries are preserved. A temporary is preserved by |
562 | pretending it was allocated at the previous nesting level. */ |
563 | |
564 | class GTY(()) temp_slot { |
565 | public: |
566 | /* Points to next temporary slot. */ |
567 | class temp_slot *next; |
568 | /* Points to previous temporary slot. */ |
569 | class temp_slot *prev; |
570 | /* The rtx to used to reference the slot. */ |
571 | rtx slot; |
572 | /* The size, in units, of the slot. */ |
573 | poly_int64 size; |
574 | /* The type of the object in the slot, or zero if it doesn't correspond |
575 | to a type. We use this to determine whether a slot can be reused. |
576 | It can be reused if objects of the type of the new slot will always |
577 | conflict with objects of the type of the old slot. */ |
578 | tree type; |
579 | /* The alignment (in bits) of the slot. */ |
580 | unsigned int align; |
581 | /* True if this temporary is currently in use. */ |
582 | bool in_use; |
583 | /* Nesting level at which this slot is being used. */ |
584 | int level; |
585 | /* The offset of the slot from the frame_pointer, including extra space |
586 | for alignment. This info is for combine_temp_slots. */ |
587 | poly_int64 base_offset; |
588 | /* The size of the slot, including extra space for alignment. This |
589 | info is for combine_temp_slots. */ |
590 | poly_int64 full_size; |
591 | }; |
592 | |
593 | /* Entry for the below hash table. */ |
594 | struct GTY((for_user)) temp_slot_address_entry { |
595 | hashval_t hash; |
596 | rtx address; |
597 | class temp_slot *temp_slot; |
598 | }; |
599 | |
600 | struct temp_address_hasher : ggc_ptr_hash<temp_slot_address_entry> |
601 | { |
602 | static hashval_t hash (temp_slot_address_entry *); |
603 | static bool equal (temp_slot_address_entry *, temp_slot_address_entry *); |
604 | }; |
605 | |
606 | /* A table of addresses that represent a stack slot. The table is a mapping |
607 | from address RTXen to a temp slot. */ |
608 | static GTY(()) hash_table<temp_address_hasher> *temp_slot_address_table; |
609 | static size_t n_temp_slots_in_use; |
610 | |
611 | /* Removes temporary slot TEMP from LIST. */ |
612 | |
613 | static void |
614 | cut_slot_from_list (class temp_slot *temp, class temp_slot **list) |
615 | { |
616 | if (temp->next) |
617 | temp->next->prev = temp->prev; |
618 | if (temp->prev) |
619 | temp->prev->next = temp->next; |
620 | else |
621 | *list = temp->next; |
622 | |
623 | temp->prev = temp->next = NULL; |
624 | } |
625 | |
626 | /* Inserts temporary slot TEMP to LIST. */ |
627 | |
628 | static void |
629 | insert_slot_to_list (class temp_slot *temp, class temp_slot **list) |
630 | { |
631 | temp->next = *list; |
632 | if (*list) |
633 | (*list)->prev = temp; |
634 | temp->prev = NULL; |
635 | *list = temp; |
636 | } |
637 | |
638 | /* Returns the list of used temp slots at LEVEL. */ |
639 | |
640 | static class temp_slot ** |
641 | temp_slots_at_level (int level) |
642 | { |
643 | if (level >= (int) vec_safe_length (used_temp_slots)) |
644 | vec_safe_grow_cleared (used_temp_slots, len: level + 1, exact: true); |
645 | |
646 | return &(*used_temp_slots)[level]; |
647 | } |
648 | |
649 | /* Returns the maximal temporary slot level. */ |
650 | |
651 | static int |
652 | max_slot_level (void) |
653 | { |
654 | if (!used_temp_slots) |
655 | return -1; |
656 | |
657 | return used_temp_slots->length () - 1; |
658 | } |
659 | |
660 | /* Moves temporary slot TEMP to LEVEL. */ |
661 | |
662 | static void |
663 | move_slot_to_level (class temp_slot *temp, int level) |
664 | { |
665 | cut_slot_from_list (temp, list: temp_slots_at_level (level: temp->level)); |
666 | insert_slot_to_list (temp, list: temp_slots_at_level (level)); |
667 | temp->level = level; |
668 | } |
669 | |
670 | /* Make temporary slot TEMP available. */ |
671 | |
672 | static void |
673 | make_slot_available (class temp_slot *temp) |
674 | { |
675 | cut_slot_from_list (temp, list: temp_slots_at_level (level: temp->level)); |
676 | insert_slot_to_list (temp, list: &avail_temp_slots); |
677 | temp->in_use = false; |
678 | temp->level = -1; |
679 | n_temp_slots_in_use--; |
680 | } |
681 | |
682 | /* Compute the hash value for an address -> temp slot mapping. |
683 | The value is cached on the mapping entry. */ |
684 | static hashval_t |
685 | temp_slot_address_compute_hash (struct temp_slot_address_entry *t) |
686 | { |
687 | int do_not_record = 0; |
688 | return hash_rtx (t->address, GET_MODE (t->address), |
689 | &do_not_record, NULL, false); |
690 | } |
691 | |
692 | /* Return the hash value for an address -> temp slot mapping. */ |
693 | hashval_t |
694 | temp_address_hasher::hash (temp_slot_address_entry *t) |
695 | { |
696 | return t->hash; |
697 | } |
698 | |
699 | /* Compare two address -> temp slot mapping entries. */ |
700 | bool |
701 | temp_address_hasher::equal (temp_slot_address_entry *t1, |
702 | temp_slot_address_entry *t2) |
703 | { |
704 | return exp_equiv_p (t1->address, t2->address, 0, true); |
705 | } |
706 | |
707 | /* Add ADDRESS as an alias of TEMP_SLOT to the addess -> temp slot mapping. */ |
708 | static void |
709 | insert_temp_slot_address (rtx address, class temp_slot *temp_slot) |
710 | { |
711 | struct temp_slot_address_entry *t = ggc_alloc<temp_slot_address_entry> (); |
712 | t->address = copy_rtx (address); |
713 | t->temp_slot = temp_slot; |
714 | t->hash = temp_slot_address_compute_hash (t); |
715 | *temp_slot_address_table->find_slot_with_hash (comparable: t, hash: t->hash, insert: INSERT) = t; |
716 | } |
717 | |
718 | /* Remove an address -> temp slot mapping entry if the temp slot is |
719 | not in use anymore. Callback for remove_unused_temp_slot_addresses. */ |
720 | int |
721 | remove_unused_temp_slot_addresses_1 (temp_slot_address_entry **slot, void *) |
722 | { |
723 | const struct temp_slot_address_entry *t = *slot; |
724 | if (! t->temp_slot->in_use) |
725 | temp_slot_address_table->clear_slot (slot); |
726 | return 1; |
727 | } |
728 | |
729 | /* Remove all mappings of addresses to unused temp slots. */ |
730 | static void |
731 | remove_unused_temp_slot_addresses (void) |
732 | { |
733 | /* Use quicker clearing if there aren't any active temp slots. */ |
734 | if (n_temp_slots_in_use) |
735 | temp_slot_address_table->traverse |
736 | <void *, remove_unused_temp_slot_addresses_1> (NULL); |
737 | else |
738 | temp_slot_address_table->empty (); |
739 | } |
740 | |
741 | /* Find the temp slot corresponding to the object at address X. */ |
742 | |
743 | static class temp_slot * |
744 | find_temp_slot_from_address (rtx x) |
745 | { |
746 | class temp_slot *p; |
747 | struct temp_slot_address_entry tmp, *t; |
748 | |
749 | /* First try the easy way: |
750 | See if X exists in the address -> temp slot mapping. */ |
751 | tmp.address = x; |
752 | tmp.temp_slot = NULL; |
753 | tmp.hash = temp_slot_address_compute_hash (t: &tmp); |
754 | t = temp_slot_address_table->find_with_hash (comparable: &tmp, hash: tmp.hash); |
755 | if (t) |
756 | return t->temp_slot; |
757 | |
758 | /* If we have a sum involving a register, see if it points to a temp |
759 | slot. */ |
760 | if (GET_CODE (x) == PLUS && REG_P (XEXP (x, 0)) |
761 | && (p = find_temp_slot_from_address (XEXP (x, 0))) != 0) |
762 | return p; |
763 | else if (GET_CODE (x) == PLUS && REG_P (XEXP (x, 1)) |
764 | && (p = find_temp_slot_from_address (XEXP (x, 1))) != 0) |
765 | return p; |
766 | |
767 | /* Last resort: Address is a virtual stack var address. */ |
768 | poly_int64 offset; |
769 | if (strip_offset (x, &offset) == virtual_stack_vars_rtx) |
770 | { |
771 | int i; |
772 | for (i = max_slot_level (); i >= 0; i--) |
773 | for (p = *temp_slots_at_level (level: i); p; p = p->next) |
774 | if (known_in_range_p (val: offset, pos: p->base_offset, size: p->full_size)) |
775 | return p; |
776 | } |
777 | |
778 | return NULL; |
779 | } |
780 | |
781 | /* Allocate a temporary stack slot and record it for possible later |
782 | reuse. |
783 | |
784 | MODE is the machine mode to be given to the returned rtx. |
785 | |
786 | SIZE is the size in units of the space required. We do no rounding here |
787 | since assign_stack_local will do any required rounding. |
788 | |
789 | TYPE is the type that will be used for the stack slot. */ |
790 | |
791 | rtx |
792 | assign_stack_temp_for_type (machine_mode mode, poly_int64 size, tree type) |
793 | { |
794 | unsigned int align; |
795 | class temp_slot *p, *best_p = 0, *selected = NULL, **pp; |
796 | rtx slot; |
797 | |
798 | gcc_assert (known_size_p (size)); |
799 | |
800 | align = get_stack_local_alignment (type, mode); |
801 | |
802 | /* Try to find an available, already-allocated temporary of the proper |
803 | mode which meets the size and alignment requirements. Choose the |
804 | smallest one with the closest alignment. |
805 | |
806 | If assign_stack_temp is called outside of the tree->rtl expansion, |
807 | we cannot reuse the stack slots (that may still refer to |
808 | VIRTUAL_STACK_VARS_REGNUM). */ |
809 | if (!virtuals_instantiated) |
810 | { |
811 | for (p = avail_temp_slots; p; p = p->next) |
812 | { |
813 | if (p->align >= align |
814 | && known_ge (p->size, size) |
815 | && GET_MODE (p->slot) == mode |
816 | && objects_must_conflict_p (p->type, type) |
817 | && (best_p == 0 |
818 | || (known_eq (best_p->size, p->size) |
819 | ? best_p->align > p->align |
820 | : known_ge (best_p->size, p->size)))) |
821 | { |
822 | if (p->align == align && known_eq (p->size, size)) |
823 | { |
824 | selected = p; |
825 | cut_slot_from_list (temp: selected, list: &avail_temp_slots); |
826 | best_p = 0; |
827 | break; |
828 | } |
829 | best_p = p; |
830 | } |
831 | } |
832 | } |
833 | |
834 | /* Make our best, if any, the one to use. */ |
835 | if (best_p) |
836 | { |
837 | selected = best_p; |
838 | cut_slot_from_list (temp: selected, list: &avail_temp_slots); |
839 | |
840 | /* If there are enough aligned bytes left over, make them into a new |
841 | temp_slot so that the extra bytes don't get wasted. Do this only |
842 | for BLKmode slots, so that we can be sure of the alignment. */ |
843 | if (GET_MODE (best_p->slot) == BLKmode) |
844 | { |
845 | int alignment = best_p->align / BITS_PER_UNIT; |
846 | poly_int64 rounded_size = aligned_upper_bound (value: size, align: alignment); |
847 | |
848 | if (known_ge (best_p->size - rounded_size, alignment)) |
849 | { |
850 | p = ggc_alloc<temp_slot> (); |
851 | p->in_use = false; |
852 | p->size = best_p->size - rounded_size; |
853 | p->base_offset = best_p->base_offset + rounded_size; |
854 | p->full_size = best_p->full_size - rounded_size; |
855 | p->slot = adjust_address_nv (best_p->slot, BLKmode, rounded_size); |
856 | p->align = best_p->align; |
857 | p->type = best_p->type; |
858 | insert_slot_to_list (temp: p, list: &avail_temp_slots); |
859 | |
860 | vec_safe_push (stack_slot_list, obj: p->slot); |
861 | |
862 | best_p->size = rounded_size; |
863 | best_p->full_size = rounded_size; |
864 | } |
865 | } |
866 | } |
867 | |
868 | /* If we still didn't find one, make a new temporary. */ |
869 | if (selected == 0) |
870 | { |
871 | poly_int64 frame_offset_old = frame_offset; |
872 | |
873 | p = ggc_alloc<temp_slot> (); |
874 | |
875 | /* We are passing an explicit alignment request to assign_stack_local. |
876 | One side effect of that is assign_stack_local will not round SIZE |
877 | to ensure the frame offset remains suitably aligned. |
878 | |
879 | So for requests which depended on the rounding of SIZE, we go ahead |
880 | and round it now. We also make sure ALIGNMENT is at least |
881 | BIGGEST_ALIGNMENT. */ |
882 | gcc_assert (mode != BLKmode || align == BIGGEST_ALIGNMENT); |
883 | p->slot = assign_stack_local_1 (mode, |
884 | size: (mode == BLKmode |
885 | ? aligned_upper_bound (value: size, |
886 | align: (int) align |
887 | / BITS_PER_UNIT) |
888 | : size), |
889 | align, kind: 0); |
890 | |
891 | p->align = align; |
892 | |
893 | /* The following slot size computation is necessary because we don't |
894 | know the actual size of the temporary slot until assign_stack_local |
895 | has performed all the frame alignment and size rounding for the |
896 | requested temporary. Note that extra space added for alignment |
897 | can be either above or below this stack slot depending on which |
898 | way the frame grows. We include the extra space if and only if it |
899 | is above this slot. */ |
900 | if (FRAME_GROWS_DOWNWARD) |
901 | p->size = frame_offset_old - frame_offset; |
902 | else |
903 | p->size = size; |
904 | |
905 | /* Now define the fields used by combine_temp_slots. */ |
906 | if (FRAME_GROWS_DOWNWARD) |
907 | { |
908 | p->base_offset = frame_offset; |
909 | p->full_size = frame_offset_old - frame_offset; |
910 | } |
911 | else |
912 | { |
913 | p->base_offset = frame_offset_old; |
914 | p->full_size = frame_offset - frame_offset_old; |
915 | } |
916 | |
917 | selected = p; |
918 | } |
919 | |
920 | p = selected; |
921 | p->in_use = true; |
922 | p->type = type; |
923 | p->level = temp_slot_level; |
924 | n_temp_slots_in_use++; |
925 | |
926 | pp = temp_slots_at_level (level: p->level); |
927 | insert_slot_to_list (temp: p, list: pp); |
928 | insert_temp_slot_address (XEXP (p->slot, 0), temp_slot: p); |
929 | |
930 | /* Create a new MEM rtx to avoid clobbering MEM flags of old slots. */ |
931 | slot = gen_rtx_MEM (mode, XEXP (p->slot, 0)); |
932 | vec_safe_push (stack_slot_list, obj: slot); |
933 | |
934 | /* If we know the alias set for the memory that will be used, use |
935 | it. If there's no TYPE, then we don't know anything about the |
936 | alias set for the memory. */ |
937 | set_mem_alias_set (slot, type ? get_alias_set (type) : 0); |
938 | set_mem_align (slot, align); |
939 | |
940 | /* If a type is specified, set the relevant flags. */ |
941 | if (type != 0) |
942 | MEM_VOLATILE_P (slot) = TYPE_VOLATILE (type); |
943 | MEM_NOTRAP_P (slot) = 1; |
944 | |
945 | return slot; |
946 | } |
947 | |
948 | /* Allocate a temporary stack slot and record it for possible later |
949 | reuse. First two arguments are same as in preceding function. */ |
950 | |
951 | rtx |
952 | assign_stack_temp (machine_mode mode, poly_int64 size) |
953 | { |
954 | return assign_stack_temp_for_type (mode, size, NULL_TREE); |
955 | } |
956 | |
957 | /* Assign a temporary. |
958 | If TYPE_OR_DECL is a decl, then we are doing it on behalf of the decl |
959 | and so that should be used in error messages. In either case, we |
960 | allocate of the given type. |
961 | MEMORY_REQUIRED is 1 if the result must be addressable stack memory; |
962 | it is 0 if a register is OK. |
963 | DONT_PROMOTE is 1 if we should not promote values in register |
964 | to wider modes. */ |
965 | |
966 | rtx |
967 | assign_temp (tree type_or_decl, int memory_required, |
968 | int dont_promote ATTRIBUTE_UNUSED) |
969 | { |
970 | tree type, decl; |
971 | machine_mode mode; |
972 | #ifdef PROMOTE_MODE |
973 | int unsignedp; |
974 | #endif |
975 | |
976 | if (DECL_P (type_or_decl)) |
977 | decl = type_or_decl, type = TREE_TYPE (decl); |
978 | else |
979 | decl = NULL, type = type_or_decl; |
980 | |
981 | mode = TYPE_MODE (type); |
982 | #ifdef PROMOTE_MODE |
983 | unsignedp = TYPE_UNSIGNED (type); |
984 | #endif |
985 | |
986 | /* Allocating temporaries of TREE_ADDRESSABLE type must be done in the front |
987 | end. See also create_tmp_var for the gimplification-time check. */ |
988 | gcc_assert (!TREE_ADDRESSABLE (type) && COMPLETE_TYPE_P (type)); |
989 | |
990 | if (mode == BLKmode || memory_required) |
991 | { |
992 | poly_int64 size; |
993 | rtx tmp; |
994 | |
995 | /* Unfortunately, we don't yet know how to allocate variable-sized |
996 | temporaries. However, sometimes we can find a fixed upper limit on |
997 | the size, so try that instead. */ |
998 | if (!poly_int_tree_p (TYPE_SIZE_UNIT (type), value: &size)) |
999 | size = max_int_size_in_bytes (type); |
1000 | |
1001 | /* Zero sized arrays are a GNU C extension. Set size to 1 to avoid |
1002 | problems with allocating the stack space. */ |
1003 | if (known_eq (size, 0)) |
1004 | size = 1; |
1005 | |
1006 | /* The size of the temporary may be too large to fit into an integer. */ |
1007 | /* ??? Not sure this should happen except for user silliness, so limit |
1008 | this to things that aren't compiler-generated temporaries. The |
1009 | rest of the time we'll die in assign_stack_temp_for_type. */ |
1010 | if (decl |
1011 | && !known_size_p (a: size) |
1012 | && TREE_CODE (TYPE_SIZE_UNIT (type)) == INTEGER_CST) |
1013 | { |
1014 | error ("size of variable %q+D is too large" , decl); |
1015 | size = 1; |
1016 | } |
1017 | |
1018 | tmp = assign_stack_temp_for_type (mode, size, type); |
1019 | return tmp; |
1020 | } |
1021 | |
1022 | #ifdef PROMOTE_MODE |
1023 | if (! dont_promote) |
1024 | mode = promote_mode (type, mode, &unsignedp); |
1025 | #endif |
1026 | |
1027 | return gen_reg_rtx (mode); |
1028 | } |
1029 | |
1030 | /* Combine temporary stack slots which are adjacent on the stack. |
1031 | |
1032 | This allows for better use of already allocated stack space. This is only |
1033 | done for BLKmode slots because we can be sure that we won't have alignment |
1034 | problems in this case. */ |
1035 | |
1036 | static void |
1037 | combine_temp_slots (void) |
1038 | { |
1039 | class temp_slot *p, *q, *next, *next_q; |
1040 | int num_slots; |
1041 | |
1042 | /* We can't combine slots, because the information about which slot |
1043 | is in which alias set will be lost. */ |
1044 | if (flag_strict_aliasing) |
1045 | return; |
1046 | |
1047 | /* If there are a lot of temp slots, don't do anything unless |
1048 | high levels of optimization. */ |
1049 | if (! flag_expensive_optimizations) |
1050 | for (p = avail_temp_slots, num_slots = 0; p; p = p->next, num_slots++) |
1051 | if (num_slots > 100 || (num_slots > 10 && optimize == 0)) |
1052 | return; |
1053 | |
1054 | for (p = avail_temp_slots; p; p = next) |
1055 | { |
1056 | int delete_p = 0; |
1057 | |
1058 | next = p->next; |
1059 | |
1060 | if (GET_MODE (p->slot) != BLKmode) |
1061 | continue; |
1062 | |
1063 | for (q = p->next; q; q = next_q) |
1064 | { |
1065 | int delete_q = 0; |
1066 | |
1067 | next_q = q->next; |
1068 | |
1069 | if (GET_MODE (q->slot) != BLKmode) |
1070 | continue; |
1071 | |
1072 | if (known_eq (p->base_offset + p->full_size, q->base_offset)) |
1073 | { |
1074 | /* Q comes after P; combine Q into P. */ |
1075 | p->size += q->size; |
1076 | p->full_size += q->full_size; |
1077 | delete_q = 1; |
1078 | } |
1079 | else if (known_eq (q->base_offset + q->full_size, p->base_offset)) |
1080 | { |
1081 | /* P comes after Q; combine P into Q. */ |
1082 | q->size += p->size; |
1083 | q->full_size += p->full_size; |
1084 | delete_p = 1; |
1085 | break; |
1086 | } |
1087 | if (delete_q) |
1088 | cut_slot_from_list (temp: q, list: &avail_temp_slots); |
1089 | } |
1090 | |
1091 | /* Either delete P or advance past it. */ |
1092 | if (delete_p) |
1093 | cut_slot_from_list (temp: p, list: &avail_temp_slots); |
1094 | } |
1095 | } |
1096 | |
1097 | /* Indicate that NEW_RTX is an alternate way of referring to the temp |
1098 | slot that previously was known by OLD_RTX. */ |
1099 | |
1100 | void |
1101 | update_temp_slot_address (rtx old_rtx, rtx new_rtx) |
1102 | { |
1103 | class temp_slot *p; |
1104 | |
1105 | if (rtx_equal_p (old_rtx, new_rtx)) |
1106 | return; |
1107 | |
1108 | p = find_temp_slot_from_address (x: old_rtx); |
1109 | |
1110 | /* If we didn't find one, see if both OLD_RTX is a PLUS. If so, and |
1111 | NEW_RTX is a register, see if one operand of the PLUS is a |
1112 | temporary location. If so, NEW_RTX points into it. Otherwise, |
1113 | if both OLD_RTX and NEW_RTX are a PLUS and if there is a register |
1114 | in common between them. If so, try a recursive call on those |
1115 | values. */ |
1116 | if (p == 0) |
1117 | { |
1118 | if (GET_CODE (old_rtx) != PLUS) |
1119 | return; |
1120 | |
1121 | if (REG_P (new_rtx)) |
1122 | { |
1123 | update_temp_slot_address (XEXP (old_rtx, 0), new_rtx); |
1124 | update_temp_slot_address (XEXP (old_rtx, 1), new_rtx); |
1125 | return; |
1126 | } |
1127 | else if (GET_CODE (new_rtx) != PLUS) |
1128 | return; |
1129 | |
1130 | if (rtx_equal_p (XEXP (old_rtx, 0), XEXP (new_rtx, 0))) |
1131 | update_temp_slot_address (XEXP (old_rtx, 1), XEXP (new_rtx, 1)); |
1132 | else if (rtx_equal_p (XEXP (old_rtx, 1), XEXP (new_rtx, 0))) |
1133 | update_temp_slot_address (XEXP (old_rtx, 0), XEXP (new_rtx, 1)); |
1134 | else if (rtx_equal_p (XEXP (old_rtx, 0), XEXP (new_rtx, 1))) |
1135 | update_temp_slot_address (XEXP (old_rtx, 1), XEXP (new_rtx, 0)); |
1136 | else if (rtx_equal_p (XEXP (old_rtx, 1), XEXP (new_rtx, 1))) |
1137 | update_temp_slot_address (XEXP (old_rtx, 0), XEXP (new_rtx, 0)); |
1138 | |
1139 | return; |
1140 | } |
1141 | |
1142 | /* Otherwise add an alias for the temp's address. */ |
1143 | insert_temp_slot_address (address: new_rtx, temp_slot: p); |
1144 | } |
1145 | |
1146 | /* If X could be a reference to a temporary slot, mark that slot as |
1147 | belonging to the to one level higher than the current level. If X |
1148 | matched one of our slots, just mark that one. Otherwise, we can't |
1149 | easily predict which it is, so upgrade all of them. |
1150 | |
1151 | This is called when an ({...}) construct occurs and a statement |
1152 | returns a value in memory. */ |
1153 | |
1154 | void |
1155 | preserve_temp_slots (rtx x) |
1156 | { |
1157 | class temp_slot *p = 0, *next; |
1158 | |
1159 | if (x == 0) |
1160 | return; |
1161 | |
1162 | /* If X is a register that is being used as a pointer, see if we have |
1163 | a temporary slot we know it points to. */ |
1164 | if (REG_P (x) && REG_POINTER (x)) |
1165 | p = find_temp_slot_from_address (x); |
1166 | |
1167 | /* If X is not in memory or is at a constant address, it cannot be in |
1168 | a temporary slot. */ |
1169 | if (p == 0 && (!MEM_P (x) || CONSTANT_P (XEXP (x, 0)))) |
1170 | return; |
1171 | |
1172 | /* First see if we can find a match. */ |
1173 | if (p == 0) |
1174 | p = find_temp_slot_from_address (XEXP (x, 0)); |
1175 | |
1176 | if (p != 0) |
1177 | { |
1178 | if (p->level == temp_slot_level) |
1179 | move_slot_to_level (temp: p, temp_slot_level - 1); |
1180 | return; |
1181 | } |
1182 | |
1183 | /* Otherwise, preserve all non-kept slots at this level. */ |
1184 | for (p = *temp_slots_at_level (temp_slot_level); p; p = next) |
1185 | { |
1186 | next = p->next; |
1187 | move_slot_to_level (temp: p, temp_slot_level - 1); |
1188 | } |
1189 | } |
1190 | |
1191 | /* Free all temporaries used so far. This is normally called at the |
1192 | end of generating code for a statement. */ |
1193 | |
1194 | void |
1195 | free_temp_slots (void) |
1196 | { |
1197 | class temp_slot *p, *next; |
1198 | bool some_available = false; |
1199 | |
1200 | for (p = *temp_slots_at_level (temp_slot_level); p; p = next) |
1201 | { |
1202 | next = p->next; |
1203 | make_slot_available (temp: p); |
1204 | some_available = true; |
1205 | } |
1206 | |
1207 | if (some_available) |
1208 | { |
1209 | remove_unused_temp_slot_addresses (); |
1210 | combine_temp_slots (); |
1211 | } |
1212 | } |
1213 | |
1214 | /* Push deeper into the nesting level for stack temporaries. */ |
1215 | |
1216 | void |
1217 | push_temp_slots (void) |
1218 | { |
1219 | temp_slot_level++; |
1220 | } |
1221 | |
1222 | /* Pop a temporary nesting level. All slots in use in the current level |
1223 | are freed. */ |
1224 | |
1225 | void |
1226 | pop_temp_slots (void) |
1227 | { |
1228 | free_temp_slots (); |
1229 | temp_slot_level--; |
1230 | } |
1231 | |
1232 | /* Initialize temporary slots. */ |
1233 | |
1234 | void |
1235 | init_temp_slots (void) |
1236 | { |
1237 | /* We have not allocated any temporaries yet. */ |
1238 | avail_temp_slots = 0; |
1239 | vec_alloc (used_temp_slots, nelems: 0); |
1240 | temp_slot_level = 0; |
1241 | n_temp_slots_in_use = 0; |
1242 | |
1243 | /* Set up the table to map addresses to temp slots. */ |
1244 | if (! temp_slot_address_table) |
1245 | temp_slot_address_table = hash_table<temp_address_hasher>::create_ggc (n: 32); |
1246 | else |
1247 | temp_slot_address_table->empty (); |
1248 | } |
1249 | |
1250 | /* Functions and data structures to keep track of the values hard regs |
1251 | had at the start of the function. */ |
1252 | |
1253 | /* Private type used by get_hard_reg_initial_reg, get_hard_reg_initial_val, |
1254 | and has_hard_reg_initial_val.. */ |
1255 | struct GTY(()) initial_value_pair { |
1256 | rtx hard_reg; |
1257 | rtx pseudo; |
1258 | }; |
1259 | /* ??? This could be a VEC but there is currently no way to define an |
1260 | opaque VEC type. This could be worked around by defining struct |
1261 | initial_value_pair in function.h. */ |
1262 | struct GTY(()) initial_value_struct { |
1263 | int num_entries; |
1264 | int max_entries; |
1265 | initial_value_pair * GTY ((length ("%h.num_entries" ))) entries; |
1266 | }; |
1267 | |
1268 | /* If a pseudo represents an initial hard reg (or expression), return |
1269 | it, else return NULL_RTX. */ |
1270 | |
1271 | rtx |
1272 | get_hard_reg_initial_reg (rtx reg) |
1273 | { |
1274 | struct initial_value_struct *ivs = crtl->hard_reg_initial_vals; |
1275 | int i; |
1276 | |
1277 | if (ivs == 0) |
1278 | return NULL_RTX; |
1279 | |
1280 | for (i = 0; i < ivs->num_entries; i++) |
1281 | if (rtx_equal_p (ivs->entries[i].pseudo, reg)) |
1282 | return ivs->entries[i].hard_reg; |
1283 | |
1284 | return NULL_RTX; |
1285 | } |
1286 | |
1287 | /* Make sure that there's a pseudo register of mode MODE that stores the |
1288 | initial value of hard register REGNO. Return an rtx for such a pseudo. */ |
1289 | |
1290 | rtx |
1291 | get_hard_reg_initial_val (machine_mode mode, unsigned int regno) |
1292 | { |
1293 | struct initial_value_struct *ivs; |
1294 | rtx rv; |
1295 | |
1296 | rv = has_hard_reg_initial_val (mode, regno); |
1297 | if (rv) |
1298 | return rv; |
1299 | |
1300 | ivs = crtl->hard_reg_initial_vals; |
1301 | if (ivs == 0) |
1302 | { |
1303 | ivs = ggc_alloc<initial_value_struct> (); |
1304 | ivs->num_entries = 0; |
1305 | ivs->max_entries = 5; |
1306 | ivs->entries = ggc_vec_alloc<initial_value_pair> (c: 5); |
1307 | crtl->hard_reg_initial_vals = ivs; |
1308 | } |
1309 | |
1310 | if (ivs->num_entries >= ivs->max_entries) |
1311 | { |
1312 | ivs->max_entries += 5; |
1313 | ivs->entries = GGC_RESIZEVEC (initial_value_pair, ivs->entries, |
1314 | ivs->max_entries); |
1315 | } |
1316 | |
1317 | ivs->entries[ivs->num_entries].hard_reg = gen_rtx_REG (mode, regno); |
1318 | ivs->entries[ivs->num_entries].pseudo = gen_reg_rtx (mode); |
1319 | |
1320 | return ivs->entries[ivs->num_entries++].pseudo; |
1321 | } |
1322 | |
1323 | /* See if get_hard_reg_initial_val has been used to create a pseudo |
1324 | for the initial value of hard register REGNO in mode MODE. Return |
1325 | the associated pseudo if so, otherwise return NULL. */ |
1326 | |
1327 | rtx |
1328 | has_hard_reg_initial_val (machine_mode mode, unsigned int regno) |
1329 | { |
1330 | struct initial_value_struct *ivs; |
1331 | int i; |
1332 | |
1333 | ivs = crtl->hard_reg_initial_vals; |
1334 | if (ivs != 0) |
1335 | for (i = 0; i < ivs->num_entries; i++) |
1336 | if (GET_MODE (ivs->entries[i].hard_reg) == mode |
1337 | && REGNO (ivs->entries[i].hard_reg) == regno) |
1338 | return ivs->entries[i].pseudo; |
1339 | |
1340 | return NULL_RTX; |
1341 | } |
1342 | |
1343 | void |
1344 | emit_initial_value_sets (void) |
1345 | { |
1346 | struct initial_value_struct *ivs = crtl->hard_reg_initial_vals; |
1347 | int i; |
1348 | rtx_insn *seq; |
1349 | |
1350 | if (ivs == 0) |
1351 | return; |
1352 | |
1353 | start_sequence (); |
1354 | for (i = 0; i < ivs->num_entries; i++) |
1355 | emit_move_insn (ivs->entries[i].pseudo, ivs->entries[i].hard_reg); |
1356 | seq = get_insns (); |
1357 | end_sequence (); |
1358 | |
1359 | emit_insn_at_entry (seq); |
1360 | } |
1361 | |
1362 | /* Return the hardreg-pseudoreg initial values pair entry I and |
1363 | TRUE if I is a valid entry, or FALSE if I is not a valid entry. */ |
1364 | bool |
1365 | initial_value_entry (int i, rtx *hreg, rtx *preg) |
1366 | { |
1367 | struct initial_value_struct *ivs = crtl->hard_reg_initial_vals; |
1368 | if (!ivs || i >= ivs->num_entries) |
1369 | return false; |
1370 | |
1371 | *hreg = ivs->entries[i].hard_reg; |
1372 | *preg = ivs->entries[i].pseudo; |
1373 | return true; |
1374 | } |
1375 | |
1376 | /* These routines are responsible for converting virtual register references |
1377 | to the actual hard register references once RTL generation is complete. |
1378 | |
1379 | The following four variables are used for communication between the |
1380 | routines. They contain the offsets of the virtual registers from their |
1381 | respective hard registers. */ |
1382 | |
1383 | static poly_int64 in_arg_offset; |
1384 | static poly_int64 var_offset; |
1385 | static poly_int64 dynamic_offset; |
1386 | static poly_int64 out_arg_offset; |
1387 | static poly_int64 cfa_offset; |
1388 | |
1389 | /* In most machines, the stack pointer register is equivalent to the bottom |
1390 | of the stack. */ |
1391 | |
1392 | #ifndef STACK_POINTER_OFFSET |
1393 | #define STACK_POINTER_OFFSET 0 |
1394 | #endif |
1395 | |
1396 | #if defined (REG_PARM_STACK_SPACE) && !defined (INCOMING_REG_PARM_STACK_SPACE) |
1397 | #define INCOMING_REG_PARM_STACK_SPACE REG_PARM_STACK_SPACE |
1398 | #endif |
1399 | |
1400 | /* If not defined, pick an appropriate default for the offset of dynamically |
1401 | allocated memory depending on the value of ACCUMULATE_OUTGOING_ARGS, |
1402 | INCOMING_REG_PARM_STACK_SPACE, and OUTGOING_REG_PARM_STACK_SPACE. */ |
1403 | |
1404 | #ifndef STACK_DYNAMIC_OFFSET |
1405 | |
1406 | /* The bottom of the stack points to the actual arguments. If |
1407 | REG_PARM_STACK_SPACE is defined, this includes the space for the register |
1408 | parameters. However, if OUTGOING_REG_PARM_STACK space is not defined, |
1409 | stack space for register parameters is not pushed by the caller, but |
1410 | rather part of the fixed stack areas and hence not included in |
1411 | `crtl->outgoing_args_size'. Nevertheless, we must allow |
1412 | for it when allocating stack dynamic objects. */ |
1413 | |
1414 | #ifdef INCOMING_REG_PARM_STACK_SPACE |
1415 | #define STACK_DYNAMIC_OFFSET(FNDECL) \ |
1416 | ((ACCUMULATE_OUTGOING_ARGS \ |
1417 | ? (crtl->outgoing_args_size \ |
1418 | + (OUTGOING_REG_PARM_STACK_SPACE ((!(FNDECL) ? NULL_TREE : TREE_TYPE (FNDECL))) ? 0 \ |
1419 | : INCOMING_REG_PARM_STACK_SPACE (FNDECL))) \ |
1420 | : 0) + (STACK_POINTER_OFFSET)) |
1421 | #else |
1422 | #define STACK_DYNAMIC_OFFSET(FNDECL) \ |
1423 | ((ACCUMULATE_OUTGOING_ARGS ? crtl->outgoing_args_size : poly_int64 (0)) \ |
1424 | + (STACK_POINTER_OFFSET)) |
1425 | #endif |
1426 | #endif |
1427 | |
1428 | |
1429 | /* Given a piece of RTX and a pointer to a HOST_WIDE_INT, if the RTX |
1430 | is a virtual register, return the equivalent hard register and set the |
1431 | offset indirectly through the pointer. Otherwise, return 0. */ |
1432 | |
1433 | static rtx |
1434 | instantiate_new_reg (rtx x, poly_int64 *poffset) |
1435 | { |
1436 | rtx new_rtx; |
1437 | poly_int64 offset; |
1438 | |
1439 | if (x == virtual_incoming_args_rtx) |
1440 | { |
1441 | if (stack_realign_drap) |
1442 | { |
1443 | /* Replace virtual_incoming_args_rtx with internal arg |
1444 | pointer if DRAP is used to realign stack. */ |
1445 | new_rtx = crtl->args.internal_arg_pointer; |
1446 | offset = 0; |
1447 | } |
1448 | else |
1449 | new_rtx = arg_pointer_rtx, offset = in_arg_offset; |
1450 | } |
1451 | else if (x == virtual_stack_vars_rtx) |
1452 | new_rtx = frame_pointer_rtx, offset = var_offset; |
1453 | else if (x == virtual_stack_dynamic_rtx) |
1454 | new_rtx = stack_pointer_rtx, offset = dynamic_offset; |
1455 | else if (x == virtual_outgoing_args_rtx) |
1456 | new_rtx = stack_pointer_rtx, offset = out_arg_offset; |
1457 | else if (x == virtual_cfa_rtx) |
1458 | { |
1459 | #ifdef FRAME_POINTER_CFA_OFFSET |
1460 | new_rtx = frame_pointer_rtx; |
1461 | #else |
1462 | new_rtx = arg_pointer_rtx; |
1463 | #endif |
1464 | offset = cfa_offset; |
1465 | } |
1466 | else if (x == virtual_preferred_stack_boundary_rtx) |
1467 | { |
1468 | new_rtx = GEN_INT (crtl->preferred_stack_boundary / BITS_PER_UNIT); |
1469 | offset = 0; |
1470 | } |
1471 | else |
1472 | return NULL_RTX; |
1473 | |
1474 | *poffset = offset; |
1475 | return new_rtx; |
1476 | } |
1477 | |
1478 | /* A subroutine of instantiate_virtual_regs. Instantiate any virtual |
1479 | registers present inside of *LOC. The expression is simplified, |
1480 | as much as possible, but is not to be considered "valid" in any sense |
1481 | implied by the target. Return true if any change is made. */ |
1482 | |
1483 | static bool |
1484 | instantiate_virtual_regs_in_rtx (rtx *loc) |
1485 | { |
1486 | if (!*loc) |
1487 | return false; |
1488 | bool changed = false; |
1489 | subrtx_ptr_iterator::array_type array; |
1490 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) |
1491 | { |
1492 | rtx *loc = *iter; |
1493 | if (rtx x = *loc) |
1494 | { |
1495 | rtx new_rtx; |
1496 | poly_int64 offset; |
1497 | switch (GET_CODE (x)) |
1498 | { |
1499 | case REG: |
1500 | new_rtx = instantiate_new_reg (x, poffset: &offset); |
1501 | if (new_rtx) |
1502 | { |
1503 | *loc = plus_constant (GET_MODE (x), new_rtx, offset); |
1504 | changed = true; |
1505 | } |
1506 | iter.skip_subrtxes (); |
1507 | break; |
1508 | |
1509 | case PLUS: |
1510 | new_rtx = instantiate_new_reg (XEXP (x, 0), poffset: &offset); |
1511 | if (new_rtx) |
1512 | { |
1513 | XEXP (x, 0) = new_rtx; |
1514 | *loc = plus_constant (GET_MODE (x), x, offset, true); |
1515 | changed = true; |
1516 | iter.skip_subrtxes (); |
1517 | break; |
1518 | } |
1519 | |
1520 | /* FIXME -- from old code */ |
1521 | /* If we have (plus (subreg (virtual-reg)) (const_int)), we know |
1522 | we can commute the PLUS and SUBREG because pointers into the |
1523 | frame are well-behaved. */ |
1524 | break; |
1525 | |
1526 | default: |
1527 | break; |
1528 | } |
1529 | } |
1530 | } |
1531 | return changed; |
1532 | } |
1533 | |
1534 | /* A subroutine of instantiate_virtual_regs_in_insn. Return true if X |
1535 | matches the predicate for insn CODE operand OPERAND. */ |
1536 | |
1537 | static bool |
1538 | safe_insn_predicate (int code, int operand, rtx x) |
1539 | { |
1540 | return code < 0 || insn_operand_matches (icode: (enum insn_code) code, opno: operand, operand: x); |
1541 | } |
1542 | |
1543 | /* A subroutine of instantiate_virtual_regs. Instantiate any virtual |
1544 | registers present inside of insn. The result will be a valid insn. */ |
1545 | |
1546 | static void |
1547 | instantiate_virtual_regs_in_insn (rtx_insn *insn) |
1548 | { |
1549 | poly_int64 offset; |
1550 | int insn_code, i; |
1551 | bool any_change = false; |
1552 | rtx set, new_rtx, x; |
1553 | rtx_insn *seq; |
1554 | |
1555 | /* There are some special cases to be handled first. */ |
1556 | set = single_set (insn); |
1557 | if (set) |
1558 | { |
1559 | /* We're allowed to assign to a virtual register. This is interpreted |
1560 | to mean that the underlying register gets assigned the inverse |
1561 | transformation. This is used, for example, in the handling of |
1562 | non-local gotos. */ |
1563 | new_rtx = instantiate_new_reg (SET_DEST (set), poffset: &offset); |
1564 | if (new_rtx) |
1565 | { |
1566 | start_sequence (); |
1567 | |
1568 | instantiate_virtual_regs_in_rtx (loc: &SET_SRC (set)); |
1569 | x = simplify_gen_binary (code: PLUS, GET_MODE (new_rtx), SET_SRC (set), |
1570 | op1: gen_int_mode (-offset, GET_MODE (new_rtx))); |
1571 | x = force_operand (x, new_rtx); |
1572 | if (x != new_rtx) |
1573 | emit_move_insn (new_rtx, x); |
1574 | |
1575 | seq = get_insns (); |
1576 | end_sequence (); |
1577 | |
1578 | emit_insn_before (seq, insn); |
1579 | delete_insn (insn); |
1580 | return; |
1581 | } |
1582 | |
1583 | /* Handle a straight copy from a virtual register by generating a |
1584 | new add insn. The difference between this and falling through |
1585 | to the generic case is avoiding a new pseudo and eliminating a |
1586 | move insn in the initial rtl stream. */ |
1587 | new_rtx = instantiate_new_reg (SET_SRC (set), poffset: &offset); |
1588 | if (new_rtx |
1589 | && maybe_ne (a: offset, b: 0) |
1590 | && REG_P (SET_DEST (set)) |
1591 | && REGNO (SET_DEST (set)) > LAST_VIRTUAL_REGISTER) |
1592 | { |
1593 | start_sequence (); |
1594 | |
1595 | x = expand_simple_binop (GET_MODE (SET_DEST (set)), PLUS, new_rtx, |
1596 | gen_int_mode (offset, |
1597 | GET_MODE (SET_DEST (set))), |
1598 | SET_DEST (set), 1, OPTAB_LIB_WIDEN); |
1599 | if (x != SET_DEST (set)) |
1600 | emit_move_insn (SET_DEST (set), x); |
1601 | |
1602 | seq = get_insns (); |
1603 | end_sequence (); |
1604 | |
1605 | emit_insn_before (seq, insn); |
1606 | delete_insn (insn); |
1607 | return; |
1608 | } |
1609 | |
1610 | extract_insn (insn); |
1611 | insn_code = INSN_CODE (insn); |
1612 | |
1613 | /* Handle a plus involving a virtual register by determining if the |
1614 | operands remain valid if they're modified in place. */ |
1615 | poly_int64 delta; |
1616 | if (GET_CODE (SET_SRC (set)) == PLUS |
1617 | && recog_data.n_operands >= 3 |
1618 | && recog_data.operand_loc[1] == &XEXP (SET_SRC (set), 0) |
1619 | && recog_data.operand_loc[2] == &XEXP (SET_SRC (set), 1) |
1620 | && poly_int_rtx_p (x: recog_data.operand[2], res: &delta) |
1621 | && (new_rtx = instantiate_new_reg (x: recog_data.operand[1], poffset: &offset))) |
1622 | { |
1623 | offset += delta; |
1624 | |
1625 | /* If the sum is zero, then replace with a plain move. */ |
1626 | if (known_eq (offset, 0) |
1627 | && REG_P (SET_DEST (set)) |
1628 | && REGNO (SET_DEST (set)) > LAST_VIRTUAL_REGISTER) |
1629 | { |
1630 | start_sequence (); |
1631 | emit_move_insn (SET_DEST (set), new_rtx); |
1632 | seq = get_insns (); |
1633 | end_sequence (); |
1634 | |
1635 | emit_insn_before (seq, insn); |
1636 | delete_insn (insn); |
1637 | return; |
1638 | } |
1639 | |
1640 | x = gen_int_mode (offset, recog_data.operand_mode[2]); |
1641 | |
1642 | /* Using validate_change and apply_change_group here leaves |
1643 | recog_data in an invalid state. Since we know exactly what |
1644 | we want to check, do those two by hand. */ |
1645 | if (safe_insn_predicate (code: insn_code, operand: 1, x: new_rtx) |
1646 | && safe_insn_predicate (code: insn_code, operand: 2, x)) |
1647 | { |
1648 | *recog_data.operand_loc[1] = recog_data.operand[1] = new_rtx; |
1649 | *recog_data.operand_loc[2] = recog_data.operand[2] = x; |
1650 | any_change = true; |
1651 | |
1652 | /* Fall through into the regular operand fixup loop in |
1653 | order to take care of operands other than 1 and 2. */ |
1654 | } |
1655 | } |
1656 | } |
1657 | else |
1658 | { |
1659 | extract_insn (insn); |
1660 | insn_code = INSN_CODE (insn); |
1661 | } |
1662 | |
1663 | /* In the general case, we expect virtual registers to appear only in |
1664 | operands, and then only as either bare registers or inside memories. */ |
1665 | for (i = 0; i < recog_data.n_operands; ++i) |
1666 | { |
1667 | x = recog_data.operand[i]; |
1668 | switch (GET_CODE (x)) |
1669 | { |
1670 | case MEM: |
1671 | { |
1672 | rtx addr = XEXP (x, 0); |
1673 | |
1674 | if (!instantiate_virtual_regs_in_rtx (loc: &addr)) |
1675 | continue; |
1676 | |
1677 | start_sequence (); |
1678 | x = replace_equiv_address (x, addr, true); |
1679 | /* It may happen that the address with the virtual reg |
1680 | was valid (e.g. based on the virtual stack reg, which might |
1681 | be acceptable to the predicates with all offsets), whereas |
1682 | the address now isn't anymore, for instance when the address |
1683 | is still offsetted, but the base reg isn't virtual-stack-reg |
1684 | anymore. Below we would do a force_reg on the whole operand, |
1685 | but this insn might actually only accept memory. Hence, |
1686 | before doing that last resort, try to reload the address into |
1687 | a register, so this operand stays a MEM. */ |
1688 | if (!safe_insn_predicate (code: insn_code, operand: i, x)) |
1689 | { |
1690 | addr = force_reg (GET_MODE (addr), addr); |
1691 | x = replace_equiv_address (x, addr, true); |
1692 | } |
1693 | seq = get_insns (); |
1694 | end_sequence (); |
1695 | if (seq) |
1696 | emit_insn_before (seq, insn); |
1697 | } |
1698 | break; |
1699 | |
1700 | case REG: |
1701 | new_rtx = instantiate_new_reg (x, poffset: &offset); |
1702 | if (new_rtx == NULL) |
1703 | continue; |
1704 | if (known_eq (offset, 0)) |
1705 | x = new_rtx; |
1706 | else |
1707 | { |
1708 | start_sequence (); |
1709 | |
1710 | /* Careful, special mode predicates may have stuff in |
1711 | insn_data[insn_code].operand[i].mode that isn't useful |
1712 | to us for computing a new value. */ |
1713 | /* ??? Recognize address_operand and/or "p" constraints |
1714 | to see if (plus new offset) is a valid before we put |
1715 | this through expand_simple_binop. */ |
1716 | x = expand_simple_binop (GET_MODE (x), PLUS, new_rtx, |
1717 | gen_int_mode (offset, GET_MODE (x)), |
1718 | NULL_RTX, 1, OPTAB_LIB_WIDEN); |
1719 | seq = get_insns (); |
1720 | end_sequence (); |
1721 | emit_insn_before (seq, insn); |
1722 | } |
1723 | break; |
1724 | |
1725 | case SUBREG: |
1726 | new_rtx = instantiate_new_reg (SUBREG_REG (x), poffset: &offset); |
1727 | if (new_rtx == NULL) |
1728 | continue; |
1729 | if (maybe_ne (a: offset, b: 0)) |
1730 | { |
1731 | start_sequence (); |
1732 | new_rtx = expand_simple_binop |
1733 | (GET_MODE (new_rtx), PLUS, new_rtx, |
1734 | gen_int_mode (offset, GET_MODE (new_rtx)), |
1735 | NULL_RTX, 1, OPTAB_LIB_WIDEN); |
1736 | seq = get_insns (); |
1737 | end_sequence (); |
1738 | emit_insn_before (seq, insn); |
1739 | } |
1740 | x = simplify_gen_subreg (outermode: recog_data.operand_mode[i], op: new_rtx, |
1741 | GET_MODE (new_rtx), SUBREG_BYTE (x)); |
1742 | gcc_assert (x); |
1743 | break; |
1744 | |
1745 | default: |
1746 | continue; |
1747 | } |
1748 | |
1749 | /* At this point, X contains the new value for the operand. |
1750 | Validate the new value vs the insn predicate. Note that |
1751 | asm insns will have insn_code -1 here. */ |
1752 | if (!safe_insn_predicate (code: insn_code, operand: i, x)) |
1753 | { |
1754 | start_sequence (); |
1755 | if (REG_P (x)) |
1756 | { |
1757 | gcc_assert (REGNO (x) <= LAST_VIRTUAL_REGISTER); |
1758 | x = copy_to_reg (x); |
1759 | } |
1760 | else |
1761 | x = force_reg (insn_data[insn_code].operand[i].mode, x); |
1762 | seq = get_insns (); |
1763 | end_sequence (); |
1764 | if (seq) |
1765 | emit_insn_before (seq, insn); |
1766 | } |
1767 | |
1768 | *recog_data.operand_loc[i] = recog_data.operand[i] = x; |
1769 | any_change = true; |
1770 | } |
1771 | |
1772 | if (any_change) |
1773 | { |
1774 | /* Propagate operand changes into the duplicates. */ |
1775 | for (i = 0; i < recog_data.n_dups; ++i) |
1776 | *recog_data.dup_loc[i] |
1777 | = copy_rtx (recog_data.operand[(unsigned)recog_data.dup_num[i]]); |
1778 | |
1779 | /* Force re-recognition of the instruction for validation. */ |
1780 | INSN_CODE (insn) = -1; |
1781 | } |
1782 | |
1783 | if (asm_noperands (PATTERN (insn)) >= 0) |
1784 | { |
1785 | if (!check_asm_operands (PATTERN (insn))) |
1786 | { |
1787 | error_for_asm (insn, "impossible constraint in %<asm%>" ); |
1788 | /* For asm goto, instead of fixing up all the edges |
1789 | just clear the template and clear input and output operands |
1790 | and strip away clobbers. */ |
1791 | if (JUMP_P (insn)) |
1792 | { |
1793 | rtx asm_op = extract_asm_operands (PATTERN (insn)); |
1794 | PATTERN (insn) = asm_op; |
1795 | PUT_MODE (x: asm_op, VOIDmode); |
1796 | ASM_OPERANDS_TEMPLATE (asm_op) = ggc_strdup ("" ); |
1797 | ASM_OPERANDS_OUTPUT_CONSTRAINT (asm_op) = "" ; |
1798 | ASM_OPERANDS_OUTPUT_IDX (asm_op) = 0; |
1799 | ASM_OPERANDS_INPUT_VEC (asm_op) = rtvec_alloc (0); |
1800 | ASM_OPERANDS_INPUT_CONSTRAINT_VEC (asm_op) = rtvec_alloc (0); |
1801 | } |
1802 | else |
1803 | delete_insn (insn); |
1804 | } |
1805 | } |
1806 | else |
1807 | { |
1808 | if (recog_memoized (insn) < 0) |
1809 | fatal_insn_not_found (insn); |
1810 | } |
1811 | } |
1812 | |
1813 | /* Subroutine of instantiate_decls. Given RTL representing a decl, |
1814 | do any instantiation required. */ |
1815 | |
1816 | void |
1817 | instantiate_decl_rtl (rtx x) |
1818 | { |
1819 | rtx addr; |
1820 | |
1821 | if (x == 0) |
1822 | return; |
1823 | |
1824 | /* If this is a CONCAT, recurse for the pieces. */ |
1825 | if (GET_CODE (x) == CONCAT) |
1826 | { |
1827 | instantiate_decl_rtl (XEXP (x, 0)); |
1828 | instantiate_decl_rtl (XEXP (x, 1)); |
1829 | return; |
1830 | } |
1831 | |
1832 | /* If this is not a MEM, no need to do anything. Similarly if the |
1833 | address is a constant or a register that is not a virtual register. */ |
1834 | if (!MEM_P (x)) |
1835 | return; |
1836 | |
1837 | addr = XEXP (x, 0); |
1838 | if (CONSTANT_P (addr) |
1839 | || (REG_P (addr) |
1840 | && !VIRTUAL_REGISTER_P (addr))) |
1841 | return; |
1842 | |
1843 | instantiate_virtual_regs_in_rtx (loc: &XEXP (x, 0)); |
1844 | } |
1845 | |
1846 | /* Helper for instantiate_decls called via walk_tree: Process all decls |
1847 | in the given DECL_VALUE_EXPR. */ |
1848 | |
1849 | static tree |
1850 | instantiate_expr (tree *tp, int *walk_subtrees, void *data ATTRIBUTE_UNUSED) |
1851 | { |
1852 | tree t = *tp; |
1853 | if (! EXPR_P (t)) |
1854 | { |
1855 | *walk_subtrees = 0; |
1856 | if (DECL_P (t)) |
1857 | { |
1858 | if (DECL_RTL_SET_P (t)) |
1859 | instantiate_decl_rtl (DECL_RTL (t)); |
1860 | if (TREE_CODE (t) == PARM_DECL && DECL_NAMELESS (t) |
1861 | && DECL_INCOMING_RTL (t)) |
1862 | instantiate_decl_rtl (DECL_INCOMING_RTL (t)); |
1863 | if ((VAR_P (t) || TREE_CODE (t) == RESULT_DECL) |
1864 | && DECL_HAS_VALUE_EXPR_P (t)) |
1865 | { |
1866 | tree v = DECL_VALUE_EXPR (t); |
1867 | walk_tree (&v, instantiate_expr, NULL, NULL); |
1868 | } |
1869 | } |
1870 | } |
1871 | return NULL; |
1872 | } |
1873 | |
1874 | /* Subroutine of instantiate_decls: Process all decls in the given |
1875 | BLOCK node and all its subblocks. */ |
1876 | |
1877 | static void |
1878 | instantiate_decls_1 (tree let) |
1879 | { |
1880 | tree t; |
1881 | |
1882 | for (t = BLOCK_VARS (let); t; t = DECL_CHAIN (t)) |
1883 | { |
1884 | if (DECL_RTL_SET_P (t)) |
1885 | instantiate_decl_rtl (DECL_RTL (t)); |
1886 | if (VAR_P (t) && DECL_HAS_VALUE_EXPR_P (t)) |
1887 | { |
1888 | tree v = DECL_VALUE_EXPR (t); |
1889 | walk_tree (&v, instantiate_expr, NULL, NULL); |
1890 | } |
1891 | } |
1892 | |
1893 | /* Process all subblocks. */ |
1894 | for (t = BLOCK_SUBBLOCKS (let); t; t = BLOCK_CHAIN (t)) |
1895 | instantiate_decls_1 (let: t); |
1896 | } |
1897 | |
1898 | /* Scan all decls in FNDECL (both variables and parameters) and instantiate |
1899 | all virtual registers in their DECL_RTL's. */ |
1900 | |
1901 | static void |
1902 | instantiate_decls (tree fndecl) |
1903 | { |
1904 | tree decl; |
1905 | unsigned ix; |
1906 | |
1907 | /* Process all parameters of the function. */ |
1908 | for (decl = DECL_ARGUMENTS (fndecl); decl; decl = DECL_CHAIN (decl)) |
1909 | { |
1910 | instantiate_decl_rtl (DECL_RTL (decl)); |
1911 | instantiate_decl_rtl (DECL_INCOMING_RTL (decl)); |
1912 | if (DECL_HAS_VALUE_EXPR_P (decl)) |
1913 | { |
1914 | tree v = DECL_VALUE_EXPR (decl); |
1915 | walk_tree (&v, instantiate_expr, NULL, NULL); |
1916 | } |
1917 | } |
1918 | |
1919 | if ((decl = DECL_RESULT (fndecl)) |
1920 | && TREE_CODE (decl) == RESULT_DECL) |
1921 | { |
1922 | if (DECL_RTL_SET_P (decl)) |
1923 | instantiate_decl_rtl (DECL_RTL (decl)); |
1924 | if (DECL_HAS_VALUE_EXPR_P (decl)) |
1925 | { |
1926 | tree v = DECL_VALUE_EXPR (decl); |
1927 | walk_tree (&v, instantiate_expr, NULL, NULL); |
1928 | } |
1929 | } |
1930 | |
1931 | /* Process the saved static chain if it exists. */ |
1932 | decl = DECL_STRUCT_FUNCTION (fndecl)->static_chain_decl; |
1933 | if (decl && DECL_HAS_VALUE_EXPR_P (decl)) |
1934 | instantiate_decl_rtl (DECL_RTL (DECL_VALUE_EXPR (decl))); |
1935 | |
1936 | /* Now process all variables defined in the function or its subblocks. */ |
1937 | if (DECL_INITIAL (fndecl)) |
1938 | instantiate_decls_1 (DECL_INITIAL (fndecl)); |
1939 | |
1940 | FOR_EACH_LOCAL_DECL (cfun, ix, decl) |
1941 | if (DECL_RTL_SET_P (decl)) |
1942 | instantiate_decl_rtl (DECL_RTL (decl)); |
1943 | vec_free (v&: cfun->local_decls); |
1944 | } |
1945 | |
1946 | /* Return the value of STACK_DYNAMIC_OFFSET for the current function. |
1947 | This is done through a function wrapper so that the macro sees a |
1948 | predictable set of included files. */ |
1949 | |
1950 | poly_int64 |
1951 | get_stack_dynamic_offset () |
1952 | { |
1953 | return STACK_DYNAMIC_OFFSET (current_function_decl); |
1954 | } |
1955 | |
1956 | /* Pass through the INSNS of function FNDECL and convert virtual register |
1957 | references to hard register references. */ |
1958 | |
1959 | static void |
1960 | instantiate_virtual_regs (void) |
1961 | { |
1962 | rtx_insn *insn; |
1963 | |
1964 | /* Compute the offsets to use for this function. */ |
1965 | in_arg_offset = FIRST_PARM_OFFSET (current_function_decl); |
1966 | var_offset = targetm.starting_frame_offset (); |
1967 | dynamic_offset = get_stack_dynamic_offset (); |
1968 | out_arg_offset = STACK_POINTER_OFFSET; |
1969 | #ifdef FRAME_POINTER_CFA_OFFSET |
1970 | cfa_offset = FRAME_POINTER_CFA_OFFSET (current_function_decl); |
1971 | #else |
1972 | cfa_offset = ARG_POINTER_CFA_OFFSET (current_function_decl); |
1973 | #endif |
1974 | |
1975 | /* Initialize recognition, indicating that volatile is OK. */ |
1976 | init_recog (); |
1977 | |
1978 | /* Scan through all the insns, instantiating every virtual register still |
1979 | present. */ |
1980 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
1981 | if (INSN_P (insn)) |
1982 | { |
1983 | /* These patterns in the instruction stream can never be recognized. |
1984 | Fortunately, they shouldn't contain virtual registers either. */ |
1985 | if (GET_CODE (PATTERN (insn)) == USE |
1986 | || GET_CODE (PATTERN (insn)) == CLOBBER |
1987 | || GET_CODE (PATTERN (insn)) == ASM_INPUT |
1988 | || DEBUG_MARKER_INSN_P (insn)) |
1989 | continue; |
1990 | else if (DEBUG_BIND_INSN_P (insn)) |
1991 | instantiate_virtual_regs_in_rtx (INSN_VAR_LOCATION_PTR (insn)); |
1992 | else |
1993 | instantiate_virtual_regs_in_insn (insn); |
1994 | |
1995 | if (insn->deleted ()) |
1996 | continue; |
1997 | |
1998 | instantiate_virtual_regs_in_rtx (loc: ®_NOTES (insn)); |
1999 | |
2000 | /* Instantiate any virtual registers in CALL_INSN_FUNCTION_USAGE. */ |
2001 | if (CALL_P (insn)) |
2002 | instantiate_virtual_regs_in_rtx (loc: &CALL_INSN_FUNCTION_USAGE (insn)); |
2003 | } |
2004 | |
2005 | /* Instantiate the virtual registers in the DECLs for debugging purposes. */ |
2006 | instantiate_decls (fndecl: current_function_decl); |
2007 | |
2008 | targetm.instantiate_decls (); |
2009 | |
2010 | /* Indicate that, from now on, assign_stack_local should use |
2011 | frame_pointer_rtx. */ |
2012 | virtuals_instantiated = 1; |
2013 | } |
2014 | |
2015 | namespace { |
2016 | |
2017 | const pass_data pass_data_instantiate_virtual_regs = |
2018 | { |
2019 | .type: RTL_PASS, /* type */ |
2020 | .name: "vregs" , /* name */ |
2021 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
2022 | .tv_id: TV_NONE, /* tv_id */ |
2023 | .properties_required: 0, /* properties_required */ |
2024 | .properties_provided: 0, /* properties_provided */ |
2025 | .properties_destroyed: 0, /* properties_destroyed */ |
2026 | .todo_flags_start: 0, /* todo_flags_start */ |
2027 | .todo_flags_finish: 0, /* todo_flags_finish */ |
2028 | }; |
2029 | |
2030 | class pass_instantiate_virtual_regs : public rtl_opt_pass |
2031 | { |
2032 | public: |
2033 | pass_instantiate_virtual_regs (gcc::context *ctxt) |
2034 | : rtl_opt_pass (pass_data_instantiate_virtual_regs, ctxt) |
2035 | {} |
2036 | |
2037 | /* opt_pass methods: */ |
2038 | unsigned int execute (function *) final override |
2039 | { |
2040 | instantiate_virtual_regs (); |
2041 | return 0; |
2042 | } |
2043 | |
2044 | }; // class pass_instantiate_virtual_regs |
2045 | |
2046 | } // anon namespace |
2047 | |
2048 | rtl_opt_pass * |
2049 | make_pass_instantiate_virtual_regs (gcc::context *ctxt) |
2050 | { |
2051 | return new pass_instantiate_virtual_regs (ctxt); |
2052 | } |
2053 | |
2054 | |
2055 | /* Return true if EXP is an aggregate type (or a value with aggregate type). |
2056 | This means a type for which function calls must pass an address to the |
2057 | function or get an address back from the function. |
2058 | EXP may be a type node or an expression (whose type is tested). */ |
2059 | |
2060 | bool |
2061 | aggregate_value_p (const_tree exp, const_tree fntype) |
2062 | { |
2063 | const_tree type = (TYPE_P (exp)) ? exp : TREE_TYPE (exp); |
2064 | int i, regno, nregs; |
2065 | rtx reg; |
2066 | |
2067 | if (fntype) |
2068 | switch (TREE_CODE (fntype)) |
2069 | { |
2070 | case CALL_EXPR: |
2071 | { |
2072 | tree fndecl = get_callee_fndecl (fntype); |
2073 | if (fndecl) |
2074 | fntype = TREE_TYPE (fndecl); |
2075 | else if (CALL_EXPR_FN (fntype)) |
2076 | fntype = TREE_TYPE (TREE_TYPE (CALL_EXPR_FN (fntype))); |
2077 | else |
2078 | /* For internal functions, assume nothing needs to be |
2079 | returned in memory. */ |
2080 | return false; |
2081 | } |
2082 | break; |
2083 | case FUNCTION_DECL: |
2084 | fntype = TREE_TYPE (fntype); |
2085 | break; |
2086 | case FUNCTION_TYPE: |
2087 | case METHOD_TYPE: |
2088 | break; |
2089 | case IDENTIFIER_NODE: |
2090 | fntype = NULL_TREE; |
2091 | break; |
2092 | default: |
2093 | /* We don't expect other tree types here. */ |
2094 | gcc_unreachable (); |
2095 | } |
2096 | |
2097 | if (VOID_TYPE_P (type)) |
2098 | return false; |
2099 | |
2100 | if (error_operand_p (t: fntype)) |
2101 | return false; |
2102 | |
2103 | /* If a record should be passed the same as its first (and only) member |
2104 | don't pass it as an aggregate. */ |
2105 | if (TREE_CODE (type) == RECORD_TYPE && TYPE_TRANSPARENT_AGGR (type)) |
2106 | return aggregate_value_p (exp: first_field (type), fntype); |
2107 | |
2108 | /* If the front end has decided that this needs to be passed by |
2109 | reference, do so. */ |
2110 | if ((TREE_CODE (exp) == PARM_DECL || TREE_CODE (exp) == RESULT_DECL) |
2111 | && DECL_BY_REFERENCE (exp)) |
2112 | return true; |
2113 | |
2114 | /* Function types that are TREE_ADDRESSABLE force return in memory. */ |
2115 | if (fntype && TREE_ADDRESSABLE (fntype)) |
2116 | return true; |
2117 | |
2118 | /* Types that are TREE_ADDRESSABLE must be constructed in memory, |
2119 | and thus can't be returned in registers. */ |
2120 | if (TREE_ADDRESSABLE (type)) |
2121 | return true; |
2122 | |
2123 | if (TYPE_EMPTY_P (type)) |
2124 | return false; |
2125 | |
2126 | if (flag_pcc_struct_return && AGGREGATE_TYPE_P (type)) |
2127 | return true; |
2128 | |
2129 | if (targetm.calls.return_in_memory (type, fntype)) |
2130 | return true; |
2131 | |
2132 | /* Make sure we have suitable call-clobbered regs to return |
2133 | the value in; if not, we must return it in memory. */ |
2134 | reg = hard_function_value (type, 0, fntype, 0); |
2135 | |
2136 | /* If we have something other than a REG (e.g. a PARALLEL), then assume |
2137 | it is OK. */ |
2138 | if (!REG_P (reg)) |
2139 | return false; |
2140 | |
2141 | /* Use the default ABI if the type of the function isn't known. |
2142 | The scheme for handling interoperability between different ABIs |
2143 | requires us to be able to tell when we're calling a function with |
2144 | a nondefault ABI. */ |
2145 | const predefined_function_abi &abi = (fntype |
2146 | ? fntype_abi (fntype) |
2147 | : default_function_abi); |
2148 | regno = REGNO (reg); |
2149 | nregs = hard_regno_nregs (regno, TYPE_MODE (type)); |
2150 | for (i = 0; i < nregs; i++) |
2151 | if (!fixed_regs[regno + i] && !abi.clobbers_full_reg_p (regno: regno + i)) |
2152 | return true; |
2153 | |
2154 | return false; |
2155 | } |
2156 | |
2157 | /* Return true if we should assign DECL a pseudo register; false if it |
2158 | should live on the local stack. */ |
2159 | |
2160 | bool |
2161 | use_register_for_decl (const_tree decl) |
2162 | { |
2163 | if (TREE_CODE (decl) == SSA_NAME) |
2164 | { |
2165 | /* We often try to use the SSA_NAME, instead of its underlying |
2166 | decl, to get type information and guide decisions, to avoid |
2167 | differences of behavior between anonymous and named |
2168 | variables, but in this one case we have to go for the actual |
2169 | variable if there is one. The main reason is that, at least |
2170 | at -O0, we want to place user variables on the stack, but we |
2171 | don't mind using pseudos for anonymous or ignored temps. |
2172 | Should we take the SSA_NAME, we'd conclude all SSA_NAMEs |
2173 | should go in pseudos, whereas their corresponding variables |
2174 | might have to go on the stack. So, disregarding the decl |
2175 | here would negatively impact debug info at -O0, enable |
2176 | coalescing between SSA_NAMEs that ought to get different |
2177 | stack/pseudo assignments, and get the incoming argument |
2178 | processing thoroughly confused by PARM_DECLs expected to live |
2179 | in stack slots but assigned to pseudos. */ |
2180 | if (!SSA_NAME_VAR (decl)) |
2181 | return TYPE_MODE (TREE_TYPE (decl)) != BLKmode |
2182 | && !(flag_float_store && FLOAT_TYPE_P (TREE_TYPE (decl))); |
2183 | |
2184 | decl = SSA_NAME_VAR (decl); |
2185 | } |
2186 | |
2187 | /* Honor volatile. */ |
2188 | if (TREE_SIDE_EFFECTS (decl)) |
2189 | return false; |
2190 | |
2191 | /* Honor addressability. */ |
2192 | if (TREE_ADDRESSABLE (decl)) |
2193 | return false; |
2194 | |
2195 | /* RESULT_DECLs are a bit special in that they're assigned without |
2196 | regard to use_register_for_decl, but we generally only store in |
2197 | them. If we coalesce their SSA NAMEs, we'd better return a |
2198 | result that matches the assignment in expand_function_start. */ |
2199 | if (TREE_CODE (decl) == RESULT_DECL) |
2200 | { |
2201 | /* If it's not an aggregate, we're going to use a REG or a |
2202 | PARALLEL containing a REG. */ |
2203 | if (!aggregate_value_p (exp: decl, fntype: current_function_decl)) |
2204 | return true; |
2205 | |
2206 | /* If expand_function_start determines the return value, we'll |
2207 | use MEM if it's not by reference. */ |
2208 | if (cfun->returns_pcc_struct |
2209 | || (targetm.calls.struct_value_rtx |
2210 | (TREE_TYPE (current_function_decl), 1))) |
2211 | return DECL_BY_REFERENCE (decl); |
2212 | |
2213 | /* Otherwise, we're taking an extra all.function_result_decl |
2214 | argument. It's set up in assign_parms_augmented_arg_list, |
2215 | under the (negated) conditions above, and then it's used to |
2216 | set up the RESULT_DECL rtl in assign_params, after looping |
2217 | over all parameters. Now, if the RESULT_DECL is not by |
2218 | reference, we'll use a MEM either way. */ |
2219 | if (!DECL_BY_REFERENCE (decl)) |
2220 | return false; |
2221 | |
2222 | /* Otherwise, if RESULT_DECL is DECL_BY_REFERENCE, it will take |
2223 | the function_result_decl's assignment. Since it's a pointer, |
2224 | we can short-circuit a number of the tests below, and we must |
2225 | duplicate them because we don't have the function_result_decl |
2226 | to test. */ |
2227 | if (!targetm.calls.allocate_stack_slots_for_args ()) |
2228 | return true; |
2229 | /* We don't set DECL_IGNORED_P for the function_result_decl. */ |
2230 | if (optimize) |
2231 | return true; |
2232 | if (cfun->tail_call_marked) |
2233 | return true; |
2234 | /* We don't set DECL_REGISTER for the function_result_decl. */ |
2235 | return false; |
2236 | } |
2237 | |
2238 | /* Only register-like things go in registers. */ |
2239 | if (DECL_MODE (decl) == BLKmode) |
2240 | return false; |
2241 | |
2242 | /* If -ffloat-store specified, don't put explicit float variables |
2243 | into registers. */ |
2244 | /* ??? This should be checked after DECL_ARTIFICIAL, but tree-ssa |
2245 | propagates values across these stores, and it probably shouldn't. */ |
2246 | if (flag_float_store && FLOAT_TYPE_P (TREE_TYPE (decl))) |
2247 | return false; |
2248 | |
2249 | if (!targetm.calls.allocate_stack_slots_for_args ()) |
2250 | return true; |
2251 | |
2252 | /* If we're not interested in tracking debugging information for |
2253 | this decl, then we can certainly put it in a register. */ |
2254 | if (DECL_IGNORED_P (decl)) |
2255 | return true; |
2256 | |
2257 | if (optimize) |
2258 | return true; |
2259 | |
2260 | /* Thunks force a tail call even at -O0 so we need to avoid creating a |
2261 | dangling reference in case the parameter is passed by reference. */ |
2262 | if (TREE_CODE (decl) == PARM_DECL && cfun->tail_call_marked) |
2263 | return true; |
2264 | |
2265 | if (!DECL_REGISTER (decl)) |
2266 | return false; |
2267 | |
2268 | /* When not optimizing, disregard register keyword for types that |
2269 | could have methods, otherwise the methods won't be callable from |
2270 | the debugger. */ |
2271 | if (RECORD_OR_UNION_TYPE_P (TREE_TYPE (decl))) |
2272 | return false; |
2273 | |
2274 | return true; |
2275 | } |
2276 | |
2277 | /* Structures to communicate between the subroutines of assign_parms. |
2278 | The first holds data persistent across all parameters, the second |
2279 | is cleared out for each parameter. */ |
2280 | |
2281 | struct assign_parm_data_all |
2282 | { |
2283 | /* When INIT_CUMULATIVE_ARGS gets revamped, allocating CUMULATIVE_ARGS |
2284 | should become a job of the target or otherwise encapsulated. */ |
2285 | CUMULATIVE_ARGS args_so_far_v; |
2286 | cumulative_args_t args_so_far; |
2287 | struct args_size stack_args_size; |
2288 | tree function_result_decl; |
2289 | tree orig_fnargs; |
2290 | rtx_insn *first_conversion_insn; |
2291 | rtx_insn *last_conversion_insn; |
2292 | HOST_WIDE_INT pretend_args_size; |
2293 | HOST_WIDE_INT ; |
2294 | int reg_parm_stack_space; |
2295 | }; |
2296 | |
2297 | struct assign_parm_data_one |
2298 | { |
2299 | tree nominal_type; |
2300 | function_arg_info arg; |
2301 | rtx entry_parm; |
2302 | rtx stack_parm; |
2303 | machine_mode nominal_mode; |
2304 | machine_mode passed_mode; |
2305 | struct locate_and_pad_arg_data locate; |
2306 | int partial; |
2307 | }; |
2308 | |
2309 | /* A subroutine of assign_parms. Initialize ALL. */ |
2310 | |
2311 | static void |
2312 | assign_parms_initialize_all (struct assign_parm_data_all *all) |
2313 | { |
2314 | tree fntype ATTRIBUTE_UNUSED; |
2315 | |
2316 | memset (s: all, c: 0, n: sizeof (*all)); |
2317 | |
2318 | fntype = TREE_TYPE (current_function_decl); |
2319 | |
2320 | #ifdef INIT_CUMULATIVE_INCOMING_ARGS |
2321 | INIT_CUMULATIVE_INCOMING_ARGS (all->args_so_far_v, fntype, NULL_RTX); |
2322 | #else |
2323 | INIT_CUMULATIVE_ARGS (all->args_so_far_v, fntype, NULL_RTX, |
2324 | current_function_decl, -1); |
2325 | #endif |
2326 | all->args_so_far = pack_cumulative_args (arg: &all->args_so_far_v); |
2327 | |
2328 | #ifdef INCOMING_REG_PARM_STACK_SPACE |
2329 | all->reg_parm_stack_space |
2330 | = INCOMING_REG_PARM_STACK_SPACE (current_function_decl); |
2331 | #endif |
2332 | } |
2333 | |
2334 | /* If ARGS contains entries with complex types, split the entry into two |
2335 | entries of the component type. Return a new list of substitutions are |
2336 | needed, else the old list. */ |
2337 | |
2338 | static void |
2339 | split_complex_args (vec<tree> *args) |
2340 | { |
2341 | unsigned i; |
2342 | tree p; |
2343 | |
2344 | FOR_EACH_VEC_ELT (*args, i, p) |
2345 | { |
2346 | tree type = TREE_TYPE (p); |
2347 | if (TREE_CODE (type) == COMPLEX_TYPE |
2348 | && targetm.calls.split_complex_arg (type)) |
2349 | { |
2350 | tree decl; |
2351 | tree subtype = TREE_TYPE (type); |
2352 | bool addressable = TREE_ADDRESSABLE (p); |
2353 | |
2354 | /* Rewrite the PARM_DECL's type with its component. */ |
2355 | p = copy_node (p); |
2356 | TREE_TYPE (p) = subtype; |
2357 | DECL_ARG_TYPE (p) = TREE_TYPE (DECL_ARG_TYPE (p)); |
2358 | SET_DECL_MODE (p, VOIDmode); |
2359 | DECL_SIZE (p) = NULL; |
2360 | DECL_SIZE_UNIT (p) = NULL; |
2361 | /* If this arg must go in memory, put it in a pseudo here. |
2362 | We can't allow it to go in memory as per normal parms, |
2363 | because the usual place might not have the imag part |
2364 | adjacent to the real part. */ |
2365 | DECL_ARTIFICIAL (p) = addressable; |
2366 | DECL_IGNORED_P (p) = addressable; |
2367 | TREE_ADDRESSABLE (p) = 0; |
2368 | layout_decl (p, 0); |
2369 | (*args)[i] = p; |
2370 | |
2371 | /* Build a second synthetic decl. */ |
2372 | decl = build_decl (EXPR_LOCATION (p), |
2373 | PARM_DECL, NULL_TREE, subtype); |
2374 | DECL_ARG_TYPE (decl) = DECL_ARG_TYPE (p); |
2375 | DECL_ARTIFICIAL (decl) = addressable; |
2376 | DECL_IGNORED_P (decl) = addressable; |
2377 | layout_decl (decl, 0); |
2378 | args->safe_insert (ix: ++i, obj: decl); |
2379 | } |
2380 | } |
2381 | } |
2382 | |
2383 | /* A subroutine of assign_parms. Adjust the parameter list to incorporate |
2384 | the hidden struct return argument, and (abi willing) complex args. |
2385 | Return the new parameter list. */ |
2386 | |
2387 | static vec<tree> |
2388 | assign_parms_augmented_arg_list (struct assign_parm_data_all *all) |
2389 | { |
2390 | tree fndecl = current_function_decl; |
2391 | tree fntype = TREE_TYPE (fndecl); |
2392 | vec<tree> fnargs = vNULL; |
2393 | tree arg; |
2394 | |
2395 | for (arg = DECL_ARGUMENTS (fndecl); arg; arg = DECL_CHAIN (arg)) |
2396 | fnargs.safe_push (obj: arg); |
2397 | |
2398 | all->orig_fnargs = DECL_ARGUMENTS (fndecl); |
2399 | |
2400 | /* If struct value address is treated as the first argument, make it so. */ |
2401 | if (aggregate_value_p (DECL_RESULT (fndecl), fntype: fndecl) |
2402 | && ! cfun->returns_pcc_struct |
2403 | && targetm.calls.struct_value_rtx (TREE_TYPE (fndecl), 1) == 0) |
2404 | { |
2405 | tree type = build_pointer_type (TREE_TYPE (fntype)); |
2406 | tree decl; |
2407 | |
2408 | decl = build_decl (DECL_SOURCE_LOCATION (fndecl), |
2409 | PARM_DECL, get_identifier (".result_ptr" ), type); |
2410 | DECL_ARG_TYPE (decl) = type; |
2411 | DECL_ARTIFICIAL (decl) = 1; |
2412 | DECL_NAMELESS (decl) = 1; |
2413 | TREE_CONSTANT (decl) = 1; |
2414 | /* We don't set DECL_IGNORED_P or DECL_REGISTER here. If this |
2415 | changes, the end of the RESULT_DECL handling block in |
2416 | use_register_for_decl must be adjusted to match. */ |
2417 | |
2418 | DECL_CHAIN (decl) = all->orig_fnargs; |
2419 | all->orig_fnargs = decl; |
2420 | fnargs.safe_insert (ix: 0, obj: decl); |
2421 | |
2422 | all->function_result_decl = decl; |
2423 | } |
2424 | |
2425 | /* If the target wants to split complex arguments into scalars, do so. */ |
2426 | if (targetm.calls.split_complex_arg) |
2427 | split_complex_args (args: &fnargs); |
2428 | |
2429 | return fnargs; |
2430 | } |
2431 | |
2432 | /* A subroutine of assign_parms. Examine PARM and pull out type and mode |
2433 | data for the parameter. Incorporate ABI specifics such as pass-by- |
2434 | reference and type promotion. */ |
2435 | |
2436 | static void |
2437 | assign_parm_find_data_types (struct assign_parm_data_all *all, tree parm, |
2438 | struct assign_parm_data_one *data) |
2439 | { |
2440 | int unsignedp; |
2441 | |
2442 | *data = assign_parm_data_one (); |
2443 | |
2444 | /* NAMED_ARG is a misnomer. We really mean 'non-variadic'. */ |
2445 | if (!cfun->stdarg) |
2446 | data->arg.named = 1; /* No variadic parms. */ |
2447 | else if (DECL_CHAIN (parm)) |
2448 | data->arg.named = 1; /* Not the last non-variadic parm. */ |
2449 | else if (targetm.calls.strict_argument_naming (all->args_so_far)) |
2450 | data->arg.named = 1; /* Only variadic ones are unnamed. */ |
2451 | else |
2452 | data->arg.named = 0; /* Treat as variadic. */ |
2453 | |
2454 | data->nominal_type = TREE_TYPE (parm); |
2455 | data->arg.type = DECL_ARG_TYPE (parm); |
2456 | |
2457 | /* Look out for errors propagating this far. Also, if the parameter's |
2458 | type is void then its value doesn't matter. */ |
2459 | if (TREE_TYPE (parm) == error_mark_node |
2460 | /* This can happen after weird syntax errors |
2461 | or if an enum type is defined among the parms. */ |
2462 | || TREE_CODE (parm) != PARM_DECL |
2463 | || data->arg.type == NULL |
2464 | || VOID_TYPE_P (data->nominal_type)) |
2465 | { |
2466 | data->nominal_type = data->arg.type = void_type_node; |
2467 | data->nominal_mode = data->passed_mode = data->arg.mode = VOIDmode; |
2468 | return; |
2469 | } |
2470 | |
2471 | /* Find mode of arg as it is passed, and mode of arg as it should be |
2472 | during execution of this function. */ |
2473 | data->passed_mode = data->arg.mode = TYPE_MODE (data->arg.type); |
2474 | data->nominal_mode = TYPE_MODE (data->nominal_type); |
2475 | |
2476 | /* If the parm is to be passed as a transparent union or record, use the |
2477 | type of the first field for the tests below. We have already verified |
2478 | that the modes are the same. */ |
2479 | if (RECORD_OR_UNION_TYPE_P (data->arg.type) |
2480 | && TYPE_TRANSPARENT_AGGR (data->arg.type)) |
2481 | data->arg.type = TREE_TYPE (first_field (data->arg.type)); |
2482 | |
2483 | /* See if this arg was passed by invisible reference. */ |
2484 | if (apply_pass_by_reference_rules (&all->args_so_far_v, data->arg)) |
2485 | { |
2486 | data->nominal_type = data->arg.type; |
2487 | data->passed_mode = data->nominal_mode = data->arg.mode; |
2488 | } |
2489 | |
2490 | /* Find mode as it is passed by the ABI. */ |
2491 | unsignedp = TYPE_UNSIGNED (data->arg.type); |
2492 | data->arg.mode |
2493 | = promote_function_mode (data->arg.type, data->arg.mode, &unsignedp, |
2494 | TREE_TYPE (current_function_decl), 0); |
2495 | } |
2496 | |
2497 | /* A subroutine of assign_parms. Invoke setup_incoming_varargs. */ |
2498 | |
2499 | static void |
2500 | assign_parms_setup_varargs (struct assign_parm_data_all *all, |
2501 | struct assign_parm_data_one *data, bool no_rtl) |
2502 | { |
2503 | int varargs_pretend_bytes = 0; |
2504 | |
2505 | function_arg_info last_named_arg = data->arg; |
2506 | last_named_arg.named = true; |
2507 | targetm.calls.setup_incoming_varargs (all->args_so_far, last_named_arg, |
2508 | &varargs_pretend_bytes, no_rtl); |
2509 | |
2510 | /* If the back-end has requested extra stack space, record how much is |
2511 | needed. Do not change pretend_args_size otherwise since it may be |
2512 | nonzero from an earlier partial argument. */ |
2513 | if (varargs_pretend_bytes > 0) |
2514 | all->pretend_args_size = varargs_pretend_bytes; |
2515 | } |
2516 | |
2517 | /* A subroutine of assign_parms. Set DATA->ENTRY_PARM corresponding to |
2518 | the incoming location of the current parameter. */ |
2519 | |
2520 | static void |
2521 | assign_parm_find_entry_rtl (struct assign_parm_data_all *all, |
2522 | struct assign_parm_data_one *data) |
2523 | { |
2524 | HOST_WIDE_INT pretend_bytes = 0; |
2525 | rtx entry_parm; |
2526 | bool in_regs; |
2527 | |
2528 | if (data->arg.mode == VOIDmode) |
2529 | { |
2530 | data->entry_parm = data->stack_parm = const0_rtx; |
2531 | return; |
2532 | } |
2533 | |
2534 | targetm.calls.warn_parameter_passing_abi (all->args_so_far, |
2535 | data->arg.type); |
2536 | |
2537 | entry_parm = targetm.calls.function_incoming_arg (all->args_so_far, |
2538 | data->arg); |
2539 | if (entry_parm == 0) |
2540 | data->arg.mode = data->passed_mode; |
2541 | |
2542 | /* Determine parm's home in the stack, in case it arrives in the stack |
2543 | or we should pretend it did. Compute the stack position and rtx where |
2544 | the argument arrives and its size. |
2545 | |
2546 | There is one complexity here: If this was a parameter that would |
2547 | have been passed in registers, but wasn't only because it is |
2548 | __builtin_va_alist, we want locate_and_pad_parm to treat it as if |
2549 | it came in a register so that REG_PARM_STACK_SPACE isn't skipped. |
2550 | In this case, we call FUNCTION_ARG with NAMED set to 1 instead of 0 |
2551 | as it was the previous time. */ |
2552 | in_regs = (entry_parm != 0); |
2553 | #ifdef STACK_PARMS_IN_REG_PARM_AREA |
2554 | in_regs = true; |
2555 | #endif |
2556 | if (!in_regs && !data->arg.named) |
2557 | { |
2558 | if (targetm.calls.pretend_outgoing_varargs_named (all->args_so_far)) |
2559 | { |
2560 | rtx tem; |
2561 | function_arg_info named_arg = data->arg; |
2562 | named_arg.named = true; |
2563 | tem = targetm.calls.function_incoming_arg (all->args_so_far, |
2564 | named_arg); |
2565 | in_regs = tem != NULL; |
2566 | } |
2567 | } |
2568 | |
2569 | /* If this parameter was passed both in registers and in the stack, use |
2570 | the copy on the stack. */ |
2571 | if (targetm.calls.must_pass_in_stack (data->arg)) |
2572 | entry_parm = 0; |
2573 | |
2574 | if (entry_parm) |
2575 | { |
2576 | int partial; |
2577 | |
2578 | partial = targetm.calls.arg_partial_bytes (all->args_so_far, data->arg); |
2579 | data->partial = partial; |
2580 | |
2581 | /* The caller might already have allocated stack space for the |
2582 | register parameters. */ |
2583 | if (partial != 0 && all->reg_parm_stack_space == 0) |
2584 | { |
2585 | /* Part of this argument is passed in registers and part |
2586 | is passed on the stack. Ask the prologue code to extend |
2587 | the stack part so that we can recreate the full value. |
2588 | |
2589 | PRETEND_BYTES is the size of the registers we need to store. |
2590 | CURRENT_FUNCTION_PRETEND_ARGS_SIZE is the amount of extra |
2591 | stack space that the prologue should allocate. |
2592 | |
2593 | Internally, gcc assumes that the argument pointer is aligned |
2594 | to STACK_BOUNDARY bits. This is used both for alignment |
2595 | optimizations (see init_emit) and to locate arguments that are |
2596 | aligned to more than PARM_BOUNDARY bits. We must preserve this |
2597 | invariant by rounding CURRENT_FUNCTION_PRETEND_ARGS_SIZE up to |
2598 | a stack boundary. */ |
2599 | |
2600 | /* We assume at most one partial arg, and it must be the first |
2601 | argument on the stack. */ |
2602 | gcc_assert (!all->extra_pretend_bytes && !all->pretend_args_size); |
2603 | |
2604 | pretend_bytes = partial; |
2605 | all->pretend_args_size = CEIL_ROUND (pretend_bytes, STACK_BYTES); |
2606 | |
2607 | /* We want to align relative to the actual stack pointer, so |
2608 | don't include this in the stack size until later. */ |
2609 | all->extra_pretend_bytes = all->pretend_args_size; |
2610 | } |
2611 | } |
2612 | |
2613 | locate_and_pad_parm (data->arg.mode, data->arg.type, in_regs, |
2614 | all->reg_parm_stack_space, |
2615 | entry_parm ? data->partial : 0, current_function_decl, |
2616 | &all->stack_args_size, &data->locate); |
2617 | |
2618 | /* Update parm_stack_boundary if this parameter is passed in the |
2619 | stack. */ |
2620 | if (!in_regs && crtl->parm_stack_boundary < data->locate.boundary) |
2621 | crtl->parm_stack_boundary = data->locate.boundary; |
2622 | |
2623 | /* Adjust offsets to include the pretend args. */ |
2624 | pretend_bytes = all->extra_pretend_bytes - pretend_bytes; |
2625 | data->locate.slot_offset.constant += pretend_bytes; |
2626 | data->locate.offset.constant += pretend_bytes; |
2627 | |
2628 | data->entry_parm = entry_parm; |
2629 | } |
2630 | |
2631 | /* A subroutine of assign_parms. If there is actually space on the stack |
2632 | for this parm, count it in stack_args_size and return true. */ |
2633 | |
2634 | static bool |
2635 | assign_parm_is_stack_parm (struct assign_parm_data_all *all, |
2636 | struct assign_parm_data_one *data) |
2637 | { |
2638 | /* Trivially true if we've no incoming register. */ |
2639 | if (data->entry_parm == NULL) |
2640 | ; |
2641 | /* Also true if we're partially in registers and partially not, |
2642 | since we've arranged to drop the entire argument on the stack. */ |
2643 | else if (data->partial != 0) |
2644 | ; |
2645 | /* Also true if the target says that it's passed in both registers |
2646 | and on the stack. */ |
2647 | else if (GET_CODE (data->entry_parm) == PARALLEL |
2648 | && XEXP (XVECEXP (data->entry_parm, 0, 0), 0) == NULL_RTX) |
2649 | ; |
2650 | /* Also true if the target says that there's stack allocated for |
2651 | all register parameters. */ |
2652 | else if (all->reg_parm_stack_space > 0) |
2653 | ; |
2654 | /* Otherwise, no, this parameter has no ABI defined stack slot. */ |
2655 | else |
2656 | return false; |
2657 | |
2658 | all->stack_args_size.constant += data->locate.size.constant; |
2659 | if (data->locate.size.var) |
2660 | ADD_PARM_SIZE (all->stack_args_size, data->locate.size.var); |
2661 | |
2662 | return true; |
2663 | } |
2664 | |
2665 | /* A subroutine of assign_parms. Given that this parameter is allocated |
2666 | stack space by the ABI, find it. */ |
2667 | |
2668 | static void |
2669 | assign_parm_find_stack_rtl (tree parm, struct assign_parm_data_one *data) |
2670 | { |
2671 | rtx offset_rtx, stack_parm; |
2672 | unsigned int align, boundary; |
2673 | |
2674 | /* If we're passing this arg using a reg, make its stack home the |
2675 | aligned stack slot. */ |
2676 | if (data->entry_parm) |
2677 | offset_rtx = ARGS_SIZE_RTX (data->locate.slot_offset); |
2678 | else |
2679 | offset_rtx = ARGS_SIZE_RTX (data->locate.offset); |
2680 | |
2681 | stack_parm = crtl->args.internal_arg_pointer; |
2682 | if (offset_rtx != const0_rtx) |
2683 | stack_parm = gen_rtx_PLUS (Pmode, stack_parm, offset_rtx); |
2684 | stack_parm = gen_rtx_MEM (data->arg.mode, stack_parm); |
2685 | |
2686 | if (!data->arg.pass_by_reference) |
2687 | { |
2688 | set_mem_attributes (stack_parm, parm, 1); |
2689 | /* set_mem_attributes could set MEM_SIZE to the passed mode's size, |
2690 | while promoted mode's size is needed. */ |
2691 | if (data->arg.mode != BLKmode |
2692 | && data->arg.mode != DECL_MODE (parm)) |
2693 | { |
2694 | set_mem_size (stack_parm, GET_MODE_SIZE (mode: data->arg.mode)); |
2695 | if (MEM_EXPR (stack_parm) && MEM_OFFSET_KNOWN_P (stack_parm)) |
2696 | { |
2697 | poly_int64 offset = subreg_lowpart_offset (DECL_MODE (parm), |
2698 | innermode: data->arg.mode); |
2699 | if (maybe_ne (a: offset, b: 0)) |
2700 | set_mem_offset (stack_parm, MEM_OFFSET (stack_parm) - offset); |
2701 | } |
2702 | } |
2703 | } |
2704 | |
2705 | boundary = data->locate.boundary; |
2706 | align = BITS_PER_UNIT; |
2707 | |
2708 | /* If we're padding upward, we know that the alignment of the slot |
2709 | is TARGET_FUNCTION_ARG_BOUNDARY. If we're using slot_offset, we're |
2710 | intentionally forcing upward padding. Otherwise we have to come |
2711 | up with a guess at the alignment based on OFFSET_RTX. */ |
2712 | poly_int64 offset; |
2713 | if (data->locate.where_pad == PAD_NONE || data->entry_parm) |
2714 | align = boundary; |
2715 | else if (data->locate.where_pad == PAD_UPWARD) |
2716 | { |
2717 | align = boundary; |
2718 | /* If the argument offset is actually more aligned than the nominal |
2719 | stack slot boundary, take advantage of that excess alignment. |
2720 | Don't make any assumptions if STACK_POINTER_OFFSET is in use. */ |
2721 | if (poly_int_rtx_p (x: offset_rtx, res: &offset) |
2722 | && known_eq (STACK_POINTER_OFFSET, 0)) |
2723 | { |
2724 | unsigned int offset_align = known_alignment (a: offset) * BITS_PER_UNIT; |
2725 | if (offset_align == 0 || offset_align > STACK_BOUNDARY) |
2726 | offset_align = STACK_BOUNDARY; |
2727 | align = MAX (align, offset_align); |
2728 | } |
2729 | } |
2730 | else if (poly_int_rtx_p (x: offset_rtx, res: &offset)) |
2731 | { |
2732 | align = least_bit_hwi (x: boundary); |
2733 | unsigned int offset_align = known_alignment (a: offset) * BITS_PER_UNIT; |
2734 | if (offset_align != 0) |
2735 | align = MIN (align, offset_align); |
2736 | } |
2737 | set_mem_align (stack_parm, align); |
2738 | |
2739 | if (data->entry_parm) |
2740 | set_reg_attrs_for_parm (data->entry_parm, stack_parm); |
2741 | |
2742 | data->stack_parm = stack_parm; |
2743 | } |
2744 | |
2745 | /* A subroutine of assign_parms. Adjust DATA->ENTRY_RTL such that it's |
2746 | always valid and contiguous. */ |
2747 | |
2748 | static void |
2749 | assign_parm_adjust_entry_rtl (struct assign_parm_data_one *data) |
2750 | { |
2751 | rtx entry_parm = data->entry_parm; |
2752 | rtx stack_parm = data->stack_parm; |
2753 | |
2754 | /* If this parm was passed part in regs and part in memory, pretend it |
2755 | arrived entirely in memory by pushing the register-part onto the stack. |
2756 | In the special case of a DImode or DFmode that is split, we could put |
2757 | it together in a pseudoreg directly, but for now that's not worth |
2758 | bothering with. */ |
2759 | if (data->partial != 0) |
2760 | { |
2761 | /* Handle calls that pass values in multiple non-contiguous |
2762 | locations. The Irix 6 ABI has examples of this. */ |
2763 | if (GET_CODE (entry_parm) == PARALLEL) |
2764 | emit_group_store (validize_mem (copy_rtx (stack_parm)), entry_parm, |
2765 | data->arg.type, int_size_in_bytes (data->arg.type)); |
2766 | else |
2767 | { |
2768 | gcc_assert (data->partial % UNITS_PER_WORD == 0); |
2769 | move_block_from_reg (REGNO (entry_parm), |
2770 | validize_mem (copy_rtx (stack_parm)), |
2771 | data->partial / UNITS_PER_WORD); |
2772 | } |
2773 | |
2774 | entry_parm = stack_parm; |
2775 | } |
2776 | |
2777 | /* If we didn't decide this parm came in a register, by default it came |
2778 | on the stack. */ |
2779 | else if (entry_parm == NULL) |
2780 | entry_parm = stack_parm; |
2781 | |
2782 | /* When an argument is passed in multiple locations, we can't make use |
2783 | of this information, but we can save some copying if the whole argument |
2784 | is passed in a single register. */ |
2785 | else if (GET_CODE (entry_parm) == PARALLEL |
2786 | && data->nominal_mode != BLKmode |
2787 | && data->passed_mode != BLKmode) |
2788 | { |
2789 | size_t i, len = XVECLEN (entry_parm, 0); |
2790 | |
2791 | for (i = 0; i < len; i++) |
2792 | if (XEXP (XVECEXP (entry_parm, 0, i), 0) != NULL_RTX |
2793 | && REG_P (XEXP (XVECEXP (entry_parm, 0, i), 0)) |
2794 | && (GET_MODE (XEXP (XVECEXP (entry_parm, 0, i), 0)) |
2795 | == data->passed_mode) |
2796 | && INTVAL (XEXP (XVECEXP (entry_parm, 0, i), 1)) == 0) |
2797 | { |
2798 | entry_parm = XEXP (XVECEXP (entry_parm, 0, i), 0); |
2799 | break; |
2800 | } |
2801 | } |
2802 | |
2803 | data->entry_parm = entry_parm; |
2804 | } |
2805 | |
2806 | /* A subroutine of assign_parms. Reconstitute any values which were |
2807 | passed in multiple registers and would fit in a single register. */ |
2808 | |
2809 | static void |
2810 | assign_parm_remove_parallels (struct assign_parm_data_one *data) |
2811 | { |
2812 | rtx entry_parm = data->entry_parm; |
2813 | |
2814 | /* Convert the PARALLEL to a REG of the same mode as the parallel. |
2815 | This can be done with register operations rather than on the |
2816 | stack, even if we will store the reconstituted parameter on the |
2817 | stack later. */ |
2818 | if (GET_CODE (entry_parm) == PARALLEL && GET_MODE (entry_parm) != BLKmode) |
2819 | { |
2820 | rtx parmreg = gen_reg_rtx (GET_MODE (entry_parm)); |
2821 | emit_group_store (parmreg, entry_parm, data->arg.type, |
2822 | GET_MODE_SIZE (GET_MODE (entry_parm))); |
2823 | entry_parm = parmreg; |
2824 | } |
2825 | |
2826 | data->entry_parm = entry_parm; |
2827 | } |
2828 | |
2829 | /* A subroutine of assign_parms. Adjust DATA->STACK_RTL such that it's |
2830 | always valid and properly aligned. */ |
2831 | |
2832 | static void |
2833 | assign_parm_adjust_stack_rtl (struct assign_parm_data_one *data) |
2834 | { |
2835 | rtx stack_parm = data->stack_parm; |
2836 | |
2837 | /* If we can't trust the parm stack slot to be aligned enough for its |
2838 | ultimate type, don't use that slot after entry. We'll make another |
2839 | stack slot, if we need one. */ |
2840 | if (stack_parm |
2841 | && ((GET_MODE_ALIGNMENT (data->nominal_mode) > MEM_ALIGN (stack_parm) |
2842 | && ((optab_handler (op: movmisalign_optab, mode: data->nominal_mode) |
2843 | != CODE_FOR_nothing) |
2844 | || targetm.slow_unaligned_access (data->nominal_mode, |
2845 | MEM_ALIGN (stack_parm)))) |
2846 | || (data->nominal_type |
2847 | && TYPE_ALIGN (data->nominal_type) > MEM_ALIGN (stack_parm) |
2848 | && MEM_ALIGN (stack_parm) < PREFERRED_STACK_BOUNDARY))) |
2849 | stack_parm = NULL; |
2850 | |
2851 | /* If parm was passed in memory, and we need to convert it on entry, |
2852 | don't store it back in that same slot. */ |
2853 | else if (data->entry_parm == stack_parm |
2854 | && data->nominal_mode != BLKmode |
2855 | && data->nominal_mode != data->passed_mode) |
2856 | stack_parm = NULL; |
2857 | |
2858 | /* If stack protection is in effect for this function, don't leave any |
2859 | pointers in their passed stack slots. */ |
2860 | else if (crtl->stack_protect_guard |
2861 | && (flag_stack_protect == SPCT_FLAG_ALL |
2862 | || data->arg.pass_by_reference |
2863 | || POINTER_TYPE_P (data->nominal_type))) |
2864 | stack_parm = NULL; |
2865 | |
2866 | data->stack_parm = stack_parm; |
2867 | } |
2868 | |
2869 | /* A subroutine of assign_parms. Return true if the current parameter |
2870 | should be stored as a BLKmode in the current frame. */ |
2871 | |
2872 | static bool |
2873 | assign_parm_setup_block_p (struct assign_parm_data_one *data) |
2874 | { |
2875 | if (data->nominal_mode == BLKmode) |
2876 | return true; |
2877 | if (GET_MODE (data->entry_parm) == BLKmode) |
2878 | return true; |
2879 | |
2880 | #ifdef BLOCK_REG_PADDING |
2881 | /* Only assign_parm_setup_block knows how to deal with register arguments |
2882 | that are padded at the least significant end. */ |
2883 | if (REG_P (data->entry_parm) |
2884 | && known_lt (GET_MODE_SIZE (data->arg.mode), UNITS_PER_WORD) |
2885 | && (BLOCK_REG_PADDING (data->passed_mode, data->arg.type, 1) |
2886 | == (BYTES_BIG_ENDIAN ? PAD_UPWARD : PAD_DOWNWARD))) |
2887 | return true; |
2888 | #endif |
2889 | |
2890 | return false; |
2891 | } |
2892 | |
2893 | /* A subroutine of assign_parms. Arrange for the parameter to be |
2894 | present and valid in DATA->STACK_RTL. */ |
2895 | |
2896 | static void |
2897 | assign_parm_setup_block (struct assign_parm_data_all *all, |
2898 | tree parm, struct assign_parm_data_one *data) |
2899 | { |
2900 | rtx entry_parm = data->entry_parm; |
2901 | rtx stack_parm = data->stack_parm; |
2902 | rtx target_reg = NULL_RTX; |
2903 | bool in_conversion_seq = false; |
2904 | HOST_WIDE_INT size; |
2905 | HOST_WIDE_INT size_stored; |
2906 | |
2907 | if (GET_CODE (entry_parm) == PARALLEL) |
2908 | entry_parm = emit_group_move_into_temps (entry_parm); |
2909 | |
2910 | /* If we want the parameter in a pseudo, don't use a stack slot. */ |
2911 | if (is_gimple_reg (parm) && use_register_for_decl (decl: parm)) |
2912 | { |
2913 | tree def = ssa_default_def (cfun, parm); |
2914 | gcc_assert (def); |
2915 | machine_mode mode = promote_ssa_mode (def, NULL); |
2916 | rtx reg = gen_reg_rtx (mode); |
2917 | if (GET_CODE (reg) != CONCAT) |
2918 | stack_parm = reg; |
2919 | else |
2920 | { |
2921 | target_reg = reg; |
2922 | /* Avoid allocating a stack slot, if there isn't one |
2923 | preallocated by the ABI. It might seem like we should |
2924 | always prefer a pseudo, but converting between |
2925 | floating-point and integer modes goes through the stack |
2926 | on various machines, so it's better to use the reserved |
2927 | stack slot than to risk wasting it and allocating more |
2928 | for the conversion. */ |
2929 | if (stack_parm == NULL_RTX) |
2930 | { |
2931 | int save = generating_concat_p; |
2932 | generating_concat_p = 0; |
2933 | stack_parm = gen_reg_rtx (mode); |
2934 | generating_concat_p = save; |
2935 | } |
2936 | } |
2937 | data->stack_parm = NULL; |
2938 | } |
2939 | |
2940 | size = int_size_in_bytes (data->arg.type); |
2941 | size_stored = CEIL_ROUND (size, UNITS_PER_WORD); |
2942 | if (stack_parm == 0) |
2943 | { |
2944 | HOST_WIDE_INT parm_align |
2945 | = (STRICT_ALIGNMENT |
2946 | ? MAX (DECL_ALIGN (parm), BITS_PER_WORD) : DECL_ALIGN (parm)); |
2947 | |
2948 | SET_DECL_ALIGN (parm, parm_align); |
2949 | if (DECL_ALIGN (parm) > MAX_SUPPORTED_STACK_ALIGNMENT) |
2950 | { |
2951 | rtx allocsize = gen_int_mode (size_stored, Pmode); |
2952 | get_dynamic_stack_size (&allocsize, 0, DECL_ALIGN (parm), NULL); |
2953 | stack_parm = assign_stack_local (BLKmode, UINTVAL (allocsize), |
2954 | MAX_SUPPORTED_STACK_ALIGNMENT); |
2955 | rtx addr = align_dynamic_address (XEXP (stack_parm, 0), |
2956 | DECL_ALIGN (parm)); |
2957 | mark_reg_pointer (addr, DECL_ALIGN (parm)); |
2958 | stack_parm = gen_rtx_MEM (GET_MODE (stack_parm), addr); |
2959 | MEM_NOTRAP_P (stack_parm) = 1; |
2960 | } |
2961 | else |
2962 | stack_parm = assign_stack_local (BLKmode, size: size_stored, |
2963 | DECL_ALIGN (parm)); |
2964 | if (known_eq (GET_MODE_SIZE (GET_MODE (entry_parm)), size)) |
2965 | PUT_MODE (x: stack_parm, GET_MODE (entry_parm)); |
2966 | set_mem_attributes (stack_parm, parm, 1); |
2967 | } |
2968 | |
2969 | /* If a BLKmode arrives in registers, copy it to a stack slot. Handle |
2970 | calls that pass values in multiple non-contiguous locations. */ |
2971 | if (REG_P (entry_parm) || GET_CODE (entry_parm) == PARALLEL) |
2972 | { |
2973 | rtx mem; |
2974 | |
2975 | /* Note that we will be storing an integral number of words. |
2976 | So we have to be careful to ensure that we allocate an |
2977 | integral number of words. We do this above when we call |
2978 | assign_stack_local if space was not allocated in the argument |
2979 | list. If it was, this will not work if PARM_BOUNDARY is not |
2980 | a multiple of BITS_PER_WORD. It isn't clear how to fix this |
2981 | if it becomes a problem. Exception is when BLKmode arrives |
2982 | with arguments not conforming to word_mode. */ |
2983 | |
2984 | if (data->stack_parm == 0) |
2985 | ; |
2986 | else if (GET_CODE (entry_parm) == PARALLEL) |
2987 | ; |
2988 | else |
2989 | gcc_assert (!size || !(PARM_BOUNDARY % BITS_PER_WORD)); |
2990 | |
2991 | mem = validize_mem (copy_rtx (stack_parm)); |
2992 | |
2993 | /* Handle values in multiple non-contiguous locations. */ |
2994 | if (GET_CODE (entry_parm) == PARALLEL && !MEM_P (mem)) |
2995 | emit_group_store (mem, entry_parm, data->arg.type, size); |
2996 | else if (GET_CODE (entry_parm) == PARALLEL) |
2997 | { |
2998 | push_to_sequence2 (all->first_conversion_insn, |
2999 | all->last_conversion_insn); |
3000 | emit_group_store (mem, entry_parm, data->arg.type, size); |
3001 | all->first_conversion_insn = get_insns (); |
3002 | all->last_conversion_insn = get_last_insn (); |
3003 | end_sequence (); |
3004 | in_conversion_seq = true; |
3005 | } |
3006 | |
3007 | else if (size == 0) |
3008 | ; |
3009 | |
3010 | /* If SIZE is that of a mode no bigger than a word, just use |
3011 | that mode's store operation. */ |
3012 | else if (size <= UNITS_PER_WORD) |
3013 | { |
3014 | unsigned int bits = size * BITS_PER_UNIT; |
3015 | machine_mode mode = int_mode_for_size (size: bits, limit: 0).else_blk (); |
3016 | |
3017 | if (mode != BLKmode |
3018 | #ifdef BLOCK_REG_PADDING |
3019 | && (size == UNITS_PER_WORD |
3020 | || (BLOCK_REG_PADDING (mode, data->arg.type, 1) |
3021 | != (BYTES_BIG_ENDIAN ? PAD_UPWARD : PAD_DOWNWARD))) |
3022 | #endif |
3023 | ) |
3024 | { |
3025 | rtx reg; |
3026 | |
3027 | /* We are really truncating a word_mode value containing |
3028 | SIZE bytes into a value of mode MODE. If such an |
3029 | operation requires no actual instructions, we can refer |
3030 | to the value directly in mode MODE, otherwise we must |
3031 | start with the register in word_mode and explicitly |
3032 | convert it. */ |
3033 | if (mode == word_mode |
3034 | || TRULY_NOOP_TRUNCATION_MODES_P (mode, word_mode)) |
3035 | reg = gen_rtx_REG (mode, REGNO (entry_parm)); |
3036 | else |
3037 | { |
3038 | reg = gen_rtx_REG (word_mode, REGNO (entry_parm)); |
3039 | reg = convert_to_mode (mode, copy_to_reg (reg), 1); |
3040 | } |
3041 | |
3042 | /* We use adjust_address to get a new MEM with the mode |
3043 | changed. adjust_address is better than change_address |
3044 | for this purpose because adjust_address does not lose |
3045 | the MEM_EXPR associated with the MEM. |
3046 | |
3047 | If the MEM_EXPR is lost, then optimizations like DSE |
3048 | assume the MEM escapes and thus is not subject to DSE. */ |
3049 | emit_move_insn (adjust_address (mem, mode, 0), reg); |
3050 | } |
3051 | |
3052 | #ifdef BLOCK_REG_PADDING |
3053 | /* Storing the register in memory as a full word, as |
3054 | move_block_from_reg below would do, and then using the |
3055 | MEM in a smaller mode, has the effect of shifting right |
3056 | if BYTES_BIG_ENDIAN. If we're bypassing memory, the |
3057 | shifting must be explicit. */ |
3058 | else if (!MEM_P (mem)) |
3059 | { |
3060 | rtx x; |
3061 | |
3062 | /* If the assert below fails, we should have taken the |
3063 | mode != BLKmode path above, unless we have downward |
3064 | padding of smaller-than-word arguments on a machine |
3065 | with little-endian bytes, which would likely require |
3066 | additional changes to work correctly. */ |
3067 | gcc_checking_assert (BYTES_BIG_ENDIAN |
3068 | && (BLOCK_REG_PADDING (mode, |
3069 | data->arg.type, 1) |
3070 | == PAD_UPWARD)); |
3071 | |
3072 | int by = (UNITS_PER_WORD - size) * BITS_PER_UNIT; |
3073 | |
3074 | x = gen_rtx_REG (word_mode, REGNO (entry_parm)); |
3075 | x = expand_shift (RSHIFT_EXPR, word_mode, x, by, |
3076 | NULL_RTX, 1); |
3077 | x = force_reg (word_mode, x); |
3078 | x = gen_lowpart_SUBREG (GET_MODE (mem), x); |
3079 | |
3080 | emit_move_insn (mem, x); |
3081 | } |
3082 | #endif |
3083 | |
3084 | /* Blocks smaller than a word on a BYTES_BIG_ENDIAN |
3085 | machine must be aligned to the left before storing |
3086 | to memory. Note that the previous test doesn't |
3087 | handle all cases (e.g. SIZE == 3). */ |
3088 | else if (size != UNITS_PER_WORD |
3089 | #ifdef BLOCK_REG_PADDING |
3090 | && (BLOCK_REG_PADDING (mode, data->arg.type, 1) |
3091 | == PAD_DOWNWARD) |
3092 | #else |
3093 | && BYTES_BIG_ENDIAN |
3094 | #endif |
3095 | ) |
3096 | { |
3097 | rtx tem, x; |
3098 | int by = (UNITS_PER_WORD - size) * BITS_PER_UNIT; |
3099 | rtx reg = gen_rtx_REG (word_mode, REGNO (entry_parm)); |
3100 | |
3101 | x = expand_shift (LSHIFT_EXPR, word_mode, reg, by, NULL_RTX, 1); |
3102 | tem = change_address (mem, word_mode, 0); |
3103 | emit_move_insn (tem, x); |
3104 | } |
3105 | else |
3106 | move_block_from_reg (REGNO (entry_parm), mem, |
3107 | size_stored / UNITS_PER_WORD); |
3108 | } |
3109 | else if (!MEM_P (mem)) |
3110 | { |
3111 | gcc_checking_assert (size > UNITS_PER_WORD); |
3112 | #ifdef BLOCK_REG_PADDING |
3113 | gcc_checking_assert (BLOCK_REG_PADDING (GET_MODE (mem), |
3114 | data->arg.type, 0) |
3115 | == PAD_UPWARD); |
3116 | #endif |
3117 | emit_move_insn (mem, entry_parm); |
3118 | } |
3119 | else |
3120 | move_block_from_reg (REGNO (entry_parm), mem, |
3121 | size_stored / UNITS_PER_WORD); |
3122 | } |
3123 | else if (data->stack_parm == 0 && !TYPE_EMPTY_P (data->arg.type)) |
3124 | { |
3125 | push_to_sequence2 (all->first_conversion_insn, all->last_conversion_insn); |
3126 | emit_block_move (stack_parm, data->entry_parm, GEN_INT (size), |
3127 | BLOCK_OP_NORMAL); |
3128 | all->first_conversion_insn = get_insns (); |
3129 | all->last_conversion_insn = get_last_insn (); |
3130 | end_sequence (); |
3131 | in_conversion_seq = true; |
3132 | } |
3133 | |
3134 | if (target_reg) |
3135 | { |
3136 | if (!in_conversion_seq) |
3137 | emit_move_insn (target_reg, stack_parm); |
3138 | else |
3139 | { |
3140 | push_to_sequence2 (all->first_conversion_insn, |
3141 | all->last_conversion_insn); |
3142 | emit_move_insn (target_reg, stack_parm); |
3143 | all->first_conversion_insn = get_insns (); |
3144 | all->last_conversion_insn = get_last_insn (); |
3145 | end_sequence (); |
3146 | } |
3147 | stack_parm = target_reg; |
3148 | } |
3149 | |
3150 | data->stack_parm = stack_parm; |
3151 | set_parm_rtl (parm, stack_parm); |
3152 | } |
3153 | |
3154 | /* A subroutine of assign_parms. Allocate a pseudo to hold the current |
3155 | parameter. Get it there. Perform all ABI specified conversions. */ |
3156 | |
3157 | static void |
3158 | assign_parm_setup_reg (struct assign_parm_data_all *all, tree parm, |
3159 | struct assign_parm_data_one *data) |
3160 | { |
3161 | rtx parmreg, validated_mem; |
3162 | rtx equiv_stack_parm; |
3163 | machine_mode promoted_nominal_mode; |
3164 | int unsignedp = TYPE_UNSIGNED (TREE_TYPE (parm)); |
3165 | bool did_conversion = false; |
3166 | bool need_conversion, moved; |
3167 | enum insn_code icode; |
3168 | rtx rtl; |
3169 | |
3170 | /* Store the parm in a pseudoregister during the function, but we may |
3171 | need to do it in a wider mode. Using 2 here makes the result |
3172 | consistent with promote_decl_mode and thus expand_expr_real_1. */ |
3173 | promoted_nominal_mode |
3174 | = promote_function_mode (data->nominal_type, data->nominal_mode, &unsignedp, |
3175 | TREE_TYPE (current_function_decl), 2); |
3176 | |
3177 | parmreg = gen_reg_rtx (promoted_nominal_mode); |
3178 | if (!DECL_ARTIFICIAL (parm)) |
3179 | mark_user_reg (parmreg); |
3180 | |
3181 | /* If this was an item that we received a pointer to, |
3182 | set rtl appropriately. */ |
3183 | if (data->arg.pass_by_reference) |
3184 | { |
3185 | rtl = gen_rtx_MEM (TYPE_MODE (TREE_TYPE (data->arg.type)), parmreg); |
3186 | set_mem_attributes (rtl, parm, 1); |
3187 | } |
3188 | else |
3189 | rtl = parmreg; |
3190 | |
3191 | assign_parm_remove_parallels (data); |
3192 | |
3193 | /* Copy the value into the register, thus bridging between |
3194 | assign_parm_find_data_types and expand_expr_real_1. */ |
3195 | |
3196 | equiv_stack_parm = data->stack_parm; |
3197 | validated_mem = validize_mem (copy_rtx (data->entry_parm)); |
3198 | |
3199 | need_conversion = (data->nominal_mode != data->passed_mode |
3200 | || promoted_nominal_mode != data->arg.mode); |
3201 | moved = false; |
3202 | |
3203 | if (need_conversion |
3204 | && GET_MODE_CLASS (data->nominal_mode) == MODE_INT |
3205 | && data->nominal_mode == data->passed_mode |
3206 | && data->nominal_mode == GET_MODE (data->entry_parm)) |
3207 | { |
3208 | /* ENTRY_PARM has been converted to PROMOTED_MODE, its |
3209 | mode, by the caller. We now have to convert it to |
3210 | NOMINAL_MODE, if different. However, PARMREG may be in |
3211 | a different mode than NOMINAL_MODE if it is being stored |
3212 | promoted. |
3213 | |
3214 | If ENTRY_PARM is a hard register, it might be in a register |
3215 | not valid for operating in its mode (e.g., an odd-numbered |
3216 | register for a DFmode). In that case, moves are the only |
3217 | thing valid, so we can't do a convert from there. This |
3218 | occurs when the calling sequence allow such misaligned |
3219 | usages. |
3220 | |
3221 | In addition, the conversion may involve a call, which could |
3222 | clobber parameters which haven't been copied to pseudo |
3223 | registers yet. |
3224 | |
3225 | First, we try to emit an insn which performs the necessary |
3226 | conversion. We verify that this insn does not clobber any |
3227 | hard registers. */ |
3228 | |
3229 | rtx op0, op1; |
3230 | |
3231 | icode = can_extend_p (promoted_nominal_mode, data->passed_mode, |
3232 | unsignedp); |
3233 | |
3234 | op0 = parmreg; |
3235 | op1 = validated_mem; |
3236 | if (icode != CODE_FOR_nothing |
3237 | && insn_operand_matches (icode, opno: 0, operand: op0) |
3238 | && insn_operand_matches (icode, opno: 1, operand: op1)) |
3239 | { |
3240 | enum rtx_code code = unsignedp ? ZERO_EXTEND : SIGN_EXTEND; |
3241 | rtx_insn *insn, *insns; |
3242 | rtx t = op1; |
3243 | HARD_REG_SET hardregs; |
3244 | |
3245 | start_sequence (); |
3246 | /* If op1 is a hard register that is likely spilled, first |
3247 | force it into a pseudo, otherwise combiner might extend |
3248 | its lifetime too much. */ |
3249 | if (GET_CODE (t) == SUBREG) |
3250 | t = SUBREG_REG (t); |
3251 | if (REG_P (t) |
3252 | && HARD_REGISTER_P (t) |
3253 | && ! TEST_HARD_REG_BIT (fixed_reg_set, REGNO (t)) |
3254 | && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (t)))) |
3255 | { |
3256 | t = gen_reg_rtx (GET_MODE (op1)); |
3257 | emit_move_insn (t, op1); |
3258 | } |
3259 | else |
3260 | t = op1; |
3261 | rtx_insn *pat = gen_extend_insn (op0, t, promoted_nominal_mode, |
3262 | data->passed_mode, unsignedp); |
3263 | emit_insn (pat); |
3264 | insns = get_insns (); |
3265 | |
3266 | moved = true; |
3267 | CLEAR_HARD_REG_SET (set&: hardregs); |
3268 | for (insn = insns; insn && moved; insn = NEXT_INSN (insn)) |
3269 | { |
3270 | if (INSN_P (insn)) |
3271 | note_stores (insn, record_hard_reg_sets, &hardregs); |
3272 | if (!hard_reg_set_empty_p (x: hardregs)) |
3273 | moved = false; |
3274 | } |
3275 | |
3276 | end_sequence (); |
3277 | |
3278 | if (moved) |
3279 | { |
3280 | emit_insn (insns); |
3281 | if (equiv_stack_parm != NULL_RTX) |
3282 | equiv_stack_parm = gen_rtx_fmt_e (code, GET_MODE (parmreg), |
3283 | equiv_stack_parm); |
3284 | } |
3285 | } |
3286 | } |
3287 | |
3288 | if (moved) |
3289 | /* Nothing to do. */ |
3290 | ; |
3291 | else if (need_conversion) |
3292 | { |
3293 | /* We did not have an insn to convert directly, or the sequence |
3294 | generated appeared unsafe. We must first copy the parm to a |
3295 | pseudo reg, and save the conversion until after all |
3296 | parameters have been moved. */ |
3297 | |
3298 | int save_tree_used; |
3299 | rtx tempreg = gen_reg_rtx (GET_MODE (data->entry_parm)); |
3300 | |
3301 | emit_move_insn (tempreg, validated_mem); |
3302 | |
3303 | push_to_sequence2 (all->first_conversion_insn, all->last_conversion_insn); |
3304 | tempreg = convert_to_mode (data->nominal_mode, tempreg, unsignedp); |
3305 | |
3306 | if (partial_subreg_p (x: tempreg) |
3307 | && GET_MODE (tempreg) == data->nominal_mode |
3308 | && REG_P (SUBREG_REG (tempreg)) |
3309 | && data->nominal_mode == data->passed_mode |
3310 | && GET_MODE (SUBREG_REG (tempreg)) == GET_MODE (data->entry_parm)) |
3311 | { |
3312 | /* The argument is already sign/zero extended, so note it |
3313 | into the subreg. */ |
3314 | SUBREG_PROMOTED_VAR_P (tempreg) = 1; |
3315 | SUBREG_PROMOTED_SET (tempreg, unsignedp); |
3316 | } |
3317 | |
3318 | /* TREE_USED gets set erroneously during expand_assignment. */ |
3319 | save_tree_used = TREE_USED (parm); |
3320 | SET_DECL_RTL (parm, rtl); |
3321 | expand_assignment (parm, make_tree (data->nominal_type, tempreg), false); |
3322 | SET_DECL_RTL (parm, NULL_RTX); |
3323 | TREE_USED (parm) = save_tree_used; |
3324 | all->first_conversion_insn = get_insns (); |
3325 | all->last_conversion_insn = get_last_insn (); |
3326 | end_sequence (); |
3327 | |
3328 | did_conversion = true; |
3329 | } |
3330 | else if (MEM_P (data->entry_parm) |
3331 | && GET_MODE_ALIGNMENT (promoted_nominal_mode) |
3332 | > MEM_ALIGN (data->entry_parm) |
3333 | && (((icode = optab_handler (op: movmisalign_optab, |
3334 | mode: promoted_nominal_mode)) |
3335 | != CODE_FOR_nothing) |
3336 | || targetm.slow_unaligned_access (promoted_nominal_mode, |
3337 | MEM_ALIGN (data->entry_parm)))) |
3338 | { |
3339 | if (icode != CODE_FOR_nothing) |
3340 | emit_insn (GEN_FCN (icode) (parmreg, validated_mem)); |
3341 | else |
3342 | rtl = parmreg = extract_bit_field (validated_mem, |
3343 | GET_MODE_BITSIZE (mode: promoted_nominal_mode), 0, |
3344 | unsignedp, parmreg, |
3345 | promoted_nominal_mode, VOIDmode, false, NULL); |
3346 | } |
3347 | else |
3348 | emit_move_insn (parmreg, validated_mem); |
3349 | |
3350 | /* If we were passed a pointer but the actual value can live in a register, |
3351 | retrieve it and use it directly. Note that we cannot use nominal_mode, |
3352 | because it will have been set to Pmode above, we must use the actual mode |
3353 | of the parameter instead. */ |
3354 | if (data->arg.pass_by_reference && TYPE_MODE (TREE_TYPE (parm)) != BLKmode) |
3355 | { |
3356 | /* Use a stack slot for debugging purposes if possible. */ |
3357 | if (use_register_for_decl (decl: parm)) |
3358 | { |
3359 | parmreg = gen_reg_rtx (TYPE_MODE (TREE_TYPE (parm))); |
3360 | mark_user_reg (parmreg); |
3361 | } |
3362 | else |
3363 | { |
3364 | int align = STACK_SLOT_ALIGNMENT (TREE_TYPE (parm), |
3365 | TYPE_MODE (TREE_TYPE (parm)), |
3366 | TYPE_ALIGN (TREE_TYPE (parm))); |
3367 | parmreg |
3368 | = assign_stack_local (TYPE_MODE (TREE_TYPE (parm)), |
3369 | size: GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (parm))), |
3370 | align); |
3371 | set_mem_attributes (parmreg, parm, 1); |
3372 | } |
3373 | |
3374 | /* We need to preserve an address based on VIRTUAL_STACK_VARS_REGNUM for |
3375 | the debug info in case it is not legitimate. */ |
3376 | if (GET_MODE (parmreg) != GET_MODE (rtl)) |
3377 | { |
3378 | rtx tempreg = gen_reg_rtx (GET_MODE (rtl)); |
3379 | int unsigned_p = TYPE_UNSIGNED (TREE_TYPE (parm)); |
3380 | |
3381 | push_to_sequence2 (all->first_conversion_insn, |
3382 | all->last_conversion_insn); |
3383 | emit_move_insn (tempreg, rtl); |
3384 | tempreg = convert_to_mode (GET_MODE (parmreg), tempreg, unsigned_p); |
3385 | emit_move_insn (MEM_P (parmreg) ? copy_rtx (parmreg) : parmreg, |
3386 | tempreg); |
3387 | all->first_conversion_insn = get_insns (); |
3388 | all->last_conversion_insn = get_last_insn (); |
3389 | end_sequence (); |
3390 | |
3391 | did_conversion = true; |
3392 | } |
3393 | else |
3394 | emit_move_insn (MEM_P (parmreg) ? copy_rtx (parmreg) : parmreg, rtl); |
3395 | |
3396 | rtl = parmreg; |
3397 | |
3398 | /* STACK_PARM is the pointer, not the parm, and PARMREG is |
3399 | now the parm. */ |
3400 | data->stack_parm = NULL; |
3401 | } |
3402 | |
3403 | set_parm_rtl (parm, rtl); |
3404 | |
3405 | /* Mark the register as eliminable if we did no conversion and it was |
3406 | copied from memory at a fixed offset, and the arg pointer was not |
3407 | copied to a pseudo-reg. If the arg pointer is a pseudo reg or the |
3408 | offset formed an invalid address, such memory-equivalences as we |
3409 | make here would screw up life analysis for it. */ |
3410 | if (data->nominal_mode == data->passed_mode |
3411 | && !did_conversion |
3412 | && data->stack_parm != 0 |
3413 | && MEM_P (data->stack_parm) |
3414 | && data->locate.offset.var == 0 |
3415 | && reg_mentioned_p (virtual_incoming_args_rtx, |
3416 | XEXP (data->stack_parm, 0))) |
3417 | { |
3418 | rtx_insn *linsn = get_last_insn (); |
3419 | rtx_insn *sinsn; |
3420 | rtx set; |
3421 | |
3422 | /* Mark complex types separately. */ |
3423 | if (GET_CODE (parmreg) == CONCAT) |
3424 | { |
3425 | scalar_mode submode = GET_MODE_INNER (GET_MODE (parmreg)); |
3426 | int regnor = REGNO (XEXP (parmreg, 0)); |
3427 | int regnoi = REGNO (XEXP (parmreg, 1)); |
3428 | rtx stackr = adjust_address_nv (data->stack_parm, submode, 0); |
3429 | rtx stacki = adjust_address_nv (data->stack_parm, submode, |
3430 | GET_MODE_SIZE (submode)); |
3431 | |
3432 | /* Scan backwards for the set of the real and |
3433 | imaginary parts. */ |
3434 | for (sinsn = linsn; sinsn != 0; |
3435 | sinsn = prev_nonnote_insn (sinsn)) |
3436 | { |
3437 | set = single_set (insn: sinsn); |
3438 | if (set == 0) |
3439 | continue; |
3440 | |
3441 | if (SET_DEST (set) == regno_reg_rtx [regnoi]) |
3442 | set_unique_reg_note (sinsn, REG_EQUIV, stacki); |
3443 | else if (SET_DEST (set) == regno_reg_rtx [regnor]) |
3444 | set_unique_reg_note (sinsn, REG_EQUIV, stackr); |
3445 | } |
3446 | } |
3447 | else |
3448 | set_dst_reg_note (linsn, REG_EQUIV, equiv_stack_parm, parmreg); |
3449 | } |
3450 | |
3451 | /* For pointer data type, suggest pointer register. */ |
3452 | if (POINTER_TYPE_P (TREE_TYPE (parm))) |
3453 | mark_reg_pointer (parmreg, |
3454 | TYPE_ALIGN (TREE_TYPE (TREE_TYPE (parm)))); |
3455 | } |
3456 | |
3457 | /* A subroutine of assign_parms. Allocate stack space to hold the current |
3458 | parameter. Get it there. Perform all ABI specified conversions. */ |
3459 | |
3460 | static void |
3461 | assign_parm_setup_stack (struct assign_parm_data_all *all, tree parm, |
3462 | struct assign_parm_data_one *data) |
3463 | { |
3464 | /* Value must be stored in the stack slot STACK_PARM during function |
3465 | execution. */ |
3466 | bool to_conversion = false; |
3467 | |
3468 | assign_parm_remove_parallels (data); |
3469 | |
3470 | if (data->arg.mode != data->nominal_mode) |
3471 | { |
3472 | /* Conversion is required. */ |
3473 | rtx tempreg = gen_reg_rtx (GET_MODE (data->entry_parm)); |
3474 | |
3475 | emit_move_insn (tempreg, validize_mem (copy_rtx (data->entry_parm))); |
3476 | |
3477 | /* Some ABIs require scalar floating point modes to be passed |
3478 | in a wider scalar integer mode. We need to explicitly |
3479 | truncate to an integer mode of the correct precision before |
3480 | using a SUBREG to reinterpret as a floating point value. */ |
3481 | if (SCALAR_FLOAT_MODE_P (data->nominal_mode) |
3482 | && SCALAR_INT_MODE_P (data->arg.mode) |
3483 | && known_lt (GET_MODE_SIZE (data->nominal_mode), |
3484 | GET_MODE_SIZE (data->arg.mode))) |
3485 | tempreg = convert_wider_int_to_float (mode: data->nominal_mode, |
3486 | imode: data->arg.mode, x: tempreg); |
3487 | |
3488 | push_to_sequence2 (all->first_conversion_insn, all->last_conversion_insn); |
3489 | to_conversion = true; |
3490 | |
3491 | data->entry_parm = convert_to_mode (data->nominal_mode, tempreg, |
3492 | TYPE_UNSIGNED (TREE_TYPE (parm))); |
3493 | |
3494 | if (data->stack_parm) |
3495 | { |
3496 | poly_int64 offset |
3497 | = subreg_lowpart_offset (outermode: data->nominal_mode, |
3498 | GET_MODE (data->stack_parm)); |
3499 | /* ??? This may need a big-endian conversion on sparc64. */ |
3500 | data->stack_parm |
3501 | = adjust_address (data->stack_parm, data->nominal_mode, 0); |
3502 | if (maybe_ne (a: offset, b: 0) && MEM_OFFSET_KNOWN_P (data->stack_parm)) |
3503 | set_mem_offset (data->stack_parm, |
3504 | MEM_OFFSET (data->stack_parm) + offset); |
3505 | } |
3506 | } |
3507 | |
3508 | if (data->entry_parm != data->stack_parm) |
3509 | { |
3510 | rtx src, dest; |
3511 | |
3512 | if (data->stack_parm == 0) |
3513 | { |
3514 | int align = STACK_SLOT_ALIGNMENT (data->arg.type, |
3515 | GET_MODE (data->entry_parm), |
3516 | TYPE_ALIGN (data->arg.type)); |
3517 | if (align < (int)GET_MODE_ALIGNMENT (GET_MODE (data->entry_parm)) |
3518 | && ((optab_handler (op: movmisalign_optab, |
3519 | GET_MODE (data->entry_parm)) |
3520 | != CODE_FOR_nothing) |
3521 | || targetm.slow_unaligned_access (GET_MODE (data->entry_parm), |
3522 | align))) |
3523 | align = GET_MODE_ALIGNMENT (GET_MODE (data->entry_parm)); |
3524 | data->stack_parm |
3525 | = assign_stack_local (GET_MODE (data->entry_parm), |
3526 | size: GET_MODE_SIZE (GET_MODE (data->entry_parm)), |
3527 | align); |
3528 | align = MEM_ALIGN (data->stack_parm); |
3529 | set_mem_attributes (data->stack_parm, parm, 1); |
3530 | set_mem_align (data->stack_parm, align); |
3531 | } |
3532 | |
3533 | dest = validize_mem (copy_rtx (data->stack_parm)); |
3534 | src = validize_mem (copy_rtx (data->entry_parm)); |
3535 | |
3536 | if (TYPE_EMPTY_P (data->arg.type)) |
3537 | /* Empty types don't really need to be copied. */; |
3538 | else if (MEM_P (src)) |
3539 | { |
3540 | /* Use a block move to handle potentially misaligned entry_parm. */ |
3541 | if (!to_conversion) |
3542 | push_to_sequence2 (all->first_conversion_insn, |
3543 | all->last_conversion_insn); |
3544 | to_conversion = true; |
3545 | |
3546 | emit_block_move (dest, src, |
3547 | GEN_INT (int_size_in_bytes (data->arg.type)), |
3548 | BLOCK_OP_NORMAL); |
3549 | } |
3550 | else |
3551 | { |
3552 | if (!REG_P (src)) |
3553 | src = force_reg (GET_MODE (src), src); |
3554 | emit_move_insn (dest, src); |
3555 | } |
3556 | } |
3557 | |
3558 | if (to_conversion) |
3559 | { |
3560 | all->first_conversion_insn = get_insns (); |
3561 | all->last_conversion_insn = get_last_insn (); |
3562 | end_sequence (); |
3563 | } |
3564 | |
3565 | set_parm_rtl (parm, data->stack_parm); |
3566 | } |
3567 | |
3568 | /* A subroutine of assign_parms. If the ABI splits complex arguments, then |
3569 | undo the frobbing that we did in assign_parms_augmented_arg_list. */ |
3570 | |
3571 | static void |
3572 | assign_parms_unsplit_complex (struct assign_parm_data_all *all, |
3573 | vec<tree> fnargs) |
3574 | { |
3575 | tree parm; |
3576 | tree orig_fnargs = all->orig_fnargs; |
3577 | unsigned i = 0; |
3578 | |
3579 | for (parm = orig_fnargs; parm; parm = TREE_CHAIN (parm), ++i) |
3580 | { |
3581 | if (TREE_CODE (TREE_TYPE (parm)) == COMPLEX_TYPE |
3582 | && targetm.calls.split_complex_arg (TREE_TYPE (parm))) |
3583 | { |
3584 | rtx tmp, real, imag; |
3585 | scalar_mode inner = GET_MODE_INNER (DECL_MODE (parm)); |
3586 | |
3587 | real = DECL_RTL (fnargs[i]); |
3588 | imag = DECL_RTL (fnargs[i + 1]); |
3589 | if (inner != GET_MODE (real)) |
3590 | { |
3591 | real = gen_lowpart_SUBREG (inner, real); |
3592 | imag = gen_lowpart_SUBREG (inner, imag); |
3593 | } |
3594 | |
3595 | if (TREE_ADDRESSABLE (parm)) |
3596 | { |
3597 | rtx rmem, imem; |
3598 | HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (parm)); |
3599 | int align = STACK_SLOT_ALIGNMENT (TREE_TYPE (parm), |
3600 | DECL_MODE (parm), |
3601 | TYPE_ALIGN (TREE_TYPE (parm))); |
3602 | |
3603 | /* split_complex_arg put the real and imag parts in |
3604 | pseudos. Move them to memory. */ |
3605 | tmp = assign_stack_local (DECL_MODE (parm), size, align); |
3606 | set_mem_attributes (tmp, parm, 1); |
3607 | rmem = adjust_address_nv (tmp, inner, 0); |
3608 | imem = adjust_address_nv (tmp, inner, GET_MODE_SIZE (inner)); |
3609 | push_to_sequence2 (all->first_conversion_insn, |
3610 | all->last_conversion_insn); |
3611 | emit_move_insn (rmem, real); |
3612 | emit_move_insn (imem, imag); |
3613 | all->first_conversion_insn = get_insns (); |
3614 | all->last_conversion_insn = get_last_insn (); |
3615 | end_sequence (); |
3616 | } |
3617 | else |
3618 | tmp = gen_rtx_CONCAT (DECL_MODE (parm), real, imag); |
3619 | set_parm_rtl (parm, tmp); |
3620 | |
3621 | real = DECL_INCOMING_RTL (fnargs[i]); |
3622 | imag = DECL_INCOMING_RTL (fnargs[i + 1]); |
3623 | if (inner != GET_MODE (real)) |
3624 | { |
3625 | real = gen_lowpart_SUBREG (inner, real); |
3626 | imag = gen_lowpart_SUBREG (inner, imag); |
3627 | } |
3628 | tmp = gen_rtx_CONCAT (DECL_MODE (parm), real, imag); |
3629 | set_decl_incoming_rtl (parm, tmp, false); |
3630 | i++; |
3631 | } |
3632 | } |
3633 | } |
3634 | |
3635 | /* Assign RTL expressions to the function's parameters. This may involve |
3636 | copying them into registers and using those registers as the DECL_RTL. */ |
3637 | |
3638 | static void |
3639 | assign_parms (tree fndecl) |
3640 | { |
3641 | struct assign_parm_data_all all; |
3642 | tree parm; |
3643 | vec<tree> fnargs; |
3644 | unsigned i; |
3645 | |
3646 | crtl->args.internal_arg_pointer |
3647 | = targetm.calls.internal_arg_pointer (); |
3648 | |
3649 | assign_parms_initialize_all (all: &all); |
3650 | fnargs = assign_parms_augmented_arg_list (all: &all); |
3651 | |
3652 | if (TYPE_NO_NAMED_ARGS_STDARG_P (TREE_TYPE (fndecl))) |
3653 | { |
3654 | struct assign_parm_data_one data = {}; |
3655 | assign_parms_setup_varargs (all: &all, data: &data, no_rtl: false); |
3656 | } |
3657 | |
3658 | FOR_EACH_VEC_ELT (fnargs, i, parm) |
3659 | { |
3660 | struct assign_parm_data_one data; |
3661 | |
3662 | /* Extract the type of PARM; adjust it according to ABI. */ |
3663 | assign_parm_find_data_types (all: &all, parm, data: &data); |
3664 | |
3665 | /* Early out for errors and void parameters. */ |
3666 | if (data.passed_mode == VOIDmode) |
3667 | { |
3668 | SET_DECL_RTL (parm, const0_rtx); |
3669 | DECL_INCOMING_RTL (parm) = DECL_RTL (parm); |
3670 | continue; |
3671 | } |
3672 | |
3673 | /* Estimate stack alignment from parameter alignment. */ |
3674 | if (SUPPORTS_STACK_ALIGNMENT) |
3675 | { |
3676 | unsigned int align |
3677 | = targetm.calls.function_arg_boundary (data.arg.mode, |
3678 | data.arg.type); |
3679 | align = MINIMUM_ALIGNMENT (data.arg.type, data.arg.mode, align); |
3680 | if (TYPE_ALIGN (data.nominal_type) > align) |
3681 | align = MINIMUM_ALIGNMENT (data.nominal_type, |
3682 | TYPE_MODE (data.nominal_type), |
3683 | TYPE_ALIGN (data.nominal_type)); |
3684 | if (crtl->stack_alignment_estimated < align) |
3685 | { |
3686 | gcc_assert (!crtl->stack_realign_processed); |
3687 | crtl->stack_alignment_estimated = align; |
3688 | } |
3689 | } |
3690 | |
3691 | /* Find out where the parameter arrives in this function. */ |
3692 | assign_parm_find_entry_rtl (all: &all, data: &data); |
3693 | |
3694 | /* Find out where stack space for this parameter might be. */ |
3695 | if (assign_parm_is_stack_parm (all: &all, data: &data)) |
3696 | { |
3697 | assign_parm_find_stack_rtl (parm, data: &data); |
3698 | assign_parm_adjust_entry_rtl (data: &data); |
3699 | /* For arguments that occupy no space in the parameter |
3700 | passing area, have non-zero size and have address taken, |
3701 | force creation of a stack slot so that they have distinct |
3702 | address from other parameters. */ |
3703 | if (TYPE_EMPTY_P (data.arg.type) |
3704 | && TREE_ADDRESSABLE (parm) |
3705 | && data.entry_parm == data.stack_parm |
3706 | && MEM_P (data.entry_parm) |
3707 | && int_size_in_bytes (data.arg.type)) |
3708 | data.stack_parm = NULL_RTX; |
3709 | } |
3710 | /* Record permanently how this parm was passed. */ |
3711 | if (data.arg.pass_by_reference) |
3712 | { |
3713 | rtx incoming_rtl |
3714 | = gen_rtx_MEM (TYPE_MODE (TREE_TYPE (data.arg.type)), |
3715 | data.entry_parm); |
3716 | set_decl_incoming_rtl (parm, incoming_rtl, true); |
3717 | } |
3718 | else |
3719 | set_decl_incoming_rtl (parm, data.entry_parm, false); |
3720 | |
3721 | assign_parm_adjust_stack_rtl (data: &data); |
3722 | |
3723 | if (assign_parm_setup_block_p (data: &data)) |
3724 | assign_parm_setup_block (all: &all, parm, data: &data); |
3725 | else if (data.arg.pass_by_reference || use_register_for_decl (decl: parm)) |
3726 | assign_parm_setup_reg (all: &all, parm, data: &data); |
3727 | else |
3728 | assign_parm_setup_stack (all: &all, parm, data: &data); |
3729 | |
3730 | if (cfun->stdarg && !DECL_CHAIN (parm)) |
3731 | assign_parms_setup_varargs (all: &all, data: &data, no_rtl: false); |
3732 | |
3733 | /* Update info on where next arg arrives in registers. */ |
3734 | targetm.calls.function_arg_advance (all.args_so_far, data.arg); |
3735 | } |
3736 | |
3737 | if (targetm.calls.split_complex_arg) |
3738 | assign_parms_unsplit_complex (all: &all, fnargs); |
3739 | |
3740 | fnargs.release (); |
3741 | |
3742 | /* Output all parameter conversion instructions (possibly including calls) |
3743 | now that all parameters have been copied out of hard registers. */ |
3744 | emit_insn (all.first_conversion_insn); |
3745 | |
3746 | /* Estimate reload stack alignment from scalar return mode. */ |
3747 | if (SUPPORTS_STACK_ALIGNMENT) |
3748 | { |
3749 | if (DECL_RESULT (fndecl)) |
3750 | { |
3751 | tree type = TREE_TYPE (DECL_RESULT (fndecl)); |
3752 | machine_mode mode = TYPE_MODE (type); |
3753 | |
3754 | if (mode != BLKmode |
3755 | && mode != VOIDmode |
3756 | && !AGGREGATE_TYPE_P (type)) |
3757 | { |
3758 | unsigned int align = GET_MODE_ALIGNMENT (mode); |
3759 | if (crtl->stack_alignment_estimated < align) |
3760 | { |
3761 | gcc_assert (!crtl->stack_realign_processed); |
3762 | crtl->stack_alignment_estimated = align; |
3763 | } |
3764 | } |
3765 | } |
3766 | } |
3767 | |
3768 | /* If we are receiving a struct value address as the first argument, set up |
3769 | the RTL for the function result. As this might require code to convert |
3770 | the transmitted address to Pmode, we do this here to ensure that possible |
3771 | preliminary conversions of the address have been emitted already. */ |
3772 | if (all.function_result_decl) |
3773 | { |
3774 | tree result = DECL_RESULT (current_function_decl); |
3775 | rtx addr = DECL_RTL (all.function_result_decl); |
3776 | rtx x; |
3777 | |
3778 | if (DECL_BY_REFERENCE (result)) |
3779 | { |
3780 | SET_DECL_VALUE_EXPR (result, all.function_result_decl); |
3781 | x = addr; |
3782 | } |
3783 | else |
3784 | { |
3785 | SET_DECL_VALUE_EXPR (result, |
3786 | build1 (INDIRECT_REF, TREE_TYPE (result), |
3787 | all.function_result_decl)); |
3788 | addr = convert_memory_address (Pmode, addr); |
3789 | x = gen_rtx_MEM (DECL_MODE (result), addr); |
3790 | set_mem_attributes (x, result, 1); |
3791 | } |
3792 | |
3793 | DECL_HAS_VALUE_EXPR_P (result) = 1; |
3794 | |
3795 | set_parm_rtl (result, x); |
3796 | } |
3797 | |
3798 | /* We have aligned all the args, so add space for the pretend args. */ |
3799 | crtl->args.pretend_args_size = all.pretend_args_size; |
3800 | all.stack_args_size.constant += all.extra_pretend_bytes; |
3801 | crtl->args.size = all.stack_args_size.constant; |
3802 | |
3803 | /* Adjust function incoming argument size for alignment and |
3804 | minimum length. */ |
3805 | |
3806 | crtl->args.size = upper_bound (crtl->args.size, b: all.reg_parm_stack_space); |
3807 | crtl->args.size = aligned_upper_bound (crtl->args.size, |
3808 | PARM_BOUNDARY / BITS_PER_UNIT); |
3809 | |
3810 | if (ARGS_GROW_DOWNWARD) |
3811 | { |
3812 | crtl->args.arg_offset_rtx |
3813 | = (all.stack_args_size.var == 0 |
3814 | ? gen_int_mode (-all.stack_args_size.constant, Pmode) |
3815 | : expand_expr (size_diffop (all.stack_args_size.var, |
3816 | size_int (-all.stack_args_size.constant)), |
3817 | NULL_RTX, VOIDmode, modifier: EXPAND_NORMAL)); |
3818 | } |
3819 | else |
3820 | crtl->args.arg_offset_rtx = ARGS_SIZE_RTX (all.stack_args_size); |
3821 | |
3822 | /* See how many bytes, if any, of its args a function should try to pop |
3823 | on return. */ |
3824 | |
3825 | crtl->args.pops_args = targetm.calls.return_pops_args (fndecl, |
3826 | TREE_TYPE (fndecl), |
3827 | crtl->args.size); |
3828 | |
3829 | /* For stdarg.h function, save info about |
3830 | regs and stack space used by the named args. */ |
3831 | |
3832 | crtl->args.info = all.args_so_far_v; |
3833 | |
3834 | /* Set the rtx used for the function return value. Put this in its |
3835 | own variable so any optimizers that need this information don't have |
3836 | to include tree.h. Do this here so it gets done when an inlined |
3837 | function gets output. */ |
3838 | |
3839 | crtl->return_rtx |
3840 | = (DECL_RTL_SET_P (DECL_RESULT (fndecl)) |
3841 | ? DECL_RTL (DECL_RESULT (fndecl)) : NULL_RTX); |
3842 | |
3843 | /* If scalar return value was computed in a pseudo-reg, or was a named |
3844 | return value that got dumped to the stack, copy that to the hard |
3845 | return register. */ |
3846 | if (DECL_RTL_SET_P (DECL_RESULT (fndecl))) |
3847 | { |
3848 | tree decl_result = DECL_RESULT (fndecl); |
3849 | rtx decl_rtl = DECL_RTL (decl_result); |
3850 | |
3851 | if (REG_P (decl_rtl) |
3852 | ? REGNO (decl_rtl) >= FIRST_PSEUDO_REGISTER |
3853 | : DECL_REGISTER (decl_result)) |
3854 | { |
3855 | rtx real_decl_rtl; |
3856 | |
3857 | /* Unless the psABI says not to. */ |
3858 | if (TYPE_EMPTY_P (TREE_TYPE (decl_result))) |
3859 | real_decl_rtl = NULL_RTX; |
3860 | else |
3861 | { |
3862 | real_decl_rtl |
3863 | = targetm.calls.function_value (TREE_TYPE (decl_result), |
3864 | fndecl, true); |
3865 | REG_FUNCTION_VALUE_P (real_decl_rtl) = 1; |
3866 | } |
3867 | /* The delay slot scheduler assumes that crtl->return_rtx |
3868 | holds the hard register containing the return value, not a |
3869 | temporary pseudo. */ |
3870 | crtl->return_rtx = real_decl_rtl; |
3871 | } |
3872 | } |
3873 | } |
3874 | |
3875 | /* Gimplify the parameter list for current_function_decl. This involves |
3876 | evaluating SAVE_EXPRs of variable sized parameters and generating code |
3877 | to implement callee-copies reference parameters. Returns a sequence of |
3878 | statements to add to the beginning of the function. */ |
3879 | |
3880 | gimple_seq |
3881 | gimplify_parameters (gimple_seq *cleanup) |
3882 | { |
3883 | struct assign_parm_data_all all; |
3884 | tree parm; |
3885 | gimple_seq stmts = NULL; |
3886 | vec<tree> fnargs; |
3887 | unsigned i; |
3888 | |
3889 | assign_parms_initialize_all (all: &all); |
3890 | fnargs = assign_parms_augmented_arg_list (all: &all); |
3891 | |
3892 | FOR_EACH_VEC_ELT (fnargs, i, parm) |
3893 | { |
3894 | struct assign_parm_data_one data; |
3895 | |
3896 | /* Extract the type of PARM; adjust it according to ABI. */ |
3897 | assign_parm_find_data_types (all: &all, parm, data: &data); |
3898 | |
3899 | /* Early out for errors and void parameters. */ |
3900 | if (data.passed_mode == VOIDmode || DECL_SIZE (parm) == NULL) |
3901 | continue; |
3902 | |
3903 | /* Update info on where next arg arrives in registers. */ |
3904 | targetm.calls.function_arg_advance (all.args_so_far, data.arg); |
3905 | |
3906 | /* ??? Once upon a time variable_size stuffed parameter list |
3907 | SAVE_EXPRs (amongst others) onto a pending sizes list. This |
3908 | turned out to be less than manageable in the gimple world. |
3909 | Now we have to hunt them down ourselves. */ |
3910 | gimplify_type_sizes (TREE_TYPE (parm), &stmts); |
3911 | |
3912 | if (TREE_CODE (DECL_SIZE_UNIT (parm)) != INTEGER_CST) |
3913 | { |
3914 | gimplify_one_sizepos (&DECL_SIZE (parm), &stmts); |
3915 | gimplify_one_sizepos (&DECL_SIZE_UNIT (parm), &stmts); |
3916 | } |
3917 | |
3918 | if (data.arg.pass_by_reference) |
3919 | { |
3920 | tree type = TREE_TYPE (data.arg.type); |
3921 | function_arg_info orig_arg (type, data.arg.named); |
3922 | if (reference_callee_copied (&all.args_so_far_v, orig_arg)) |
3923 | { |
3924 | tree local, t; |
3925 | |
3926 | /* For constant-sized objects, this is trivial; for |
3927 | variable-sized objects, we have to play games. */ |
3928 | if (TREE_CODE (DECL_SIZE_UNIT (parm)) == INTEGER_CST |
3929 | && !(flag_stack_check == GENERIC_STACK_CHECK |
3930 | && compare_tree_int (DECL_SIZE_UNIT (parm), |
3931 | STACK_CHECK_MAX_VAR_SIZE) > 0)) |
3932 | { |
3933 | local = create_tmp_var (type, get_name (parm)); |
3934 | DECL_IGNORED_P (local) = 0; |
3935 | /* If PARM was addressable, move that flag over |
3936 | to the local copy, as its address will be taken, |
3937 | not the PARMs. Keep the parms address taken |
3938 | as we'll query that flag during gimplification. */ |
3939 | if (TREE_ADDRESSABLE (parm)) |
3940 | TREE_ADDRESSABLE (local) = 1; |
3941 | if (DECL_NOT_GIMPLE_REG_P (parm)) |
3942 | DECL_NOT_GIMPLE_REG_P (local) = 1; |
3943 | |
3944 | if (!is_gimple_reg (local) |
3945 | && flag_stack_reuse != SR_NONE) |
3946 | { |
3947 | tree clobber = build_clobber (type); |
3948 | gimple *clobber_stmt; |
3949 | clobber_stmt = gimple_build_assign (local, clobber); |
3950 | gimple_seq_add_stmt (cleanup, clobber_stmt); |
3951 | } |
3952 | } |
3953 | else |
3954 | { |
3955 | tree ptr_type, addr; |
3956 | |
3957 | ptr_type = build_pointer_type (type); |
3958 | addr = create_tmp_reg (ptr_type, get_name (parm)); |
3959 | DECL_IGNORED_P (addr) = 0; |
3960 | local = build_fold_indirect_ref (addr); |
3961 | |
3962 | t = build_alloca_call_expr (DECL_SIZE_UNIT (parm), |
3963 | DECL_ALIGN (parm), |
3964 | max_int_size_in_bytes (type)); |
3965 | /* The call has been built for a variable-sized object. */ |
3966 | CALL_ALLOCA_FOR_VAR_P (t) = 1; |
3967 | t = fold_convert (ptr_type, t); |
3968 | t = build2 (MODIFY_EXPR, TREE_TYPE (addr), addr, t); |
3969 | gimplify_and_add (t, &stmts); |
3970 | } |
3971 | |
3972 | gimplify_assign (local, parm, &stmts); |
3973 | |
3974 | SET_DECL_VALUE_EXPR (parm, local); |
3975 | DECL_HAS_VALUE_EXPR_P (parm) = 1; |
3976 | } |
3977 | } |
3978 | } |
3979 | |
3980 | fnargs.release (); |
3981 | |
3982 | return stmts; |
3983 | } |
3984 | |
3985 | /* Compute the size and offset from the start of the stacked arguments for a |
3986 | parm passed in mode PASSED_MODE and with type TYPE. |
3987 | |
3988 | INITIAL_OFFSET_PTR points to the current offset into the stacked |
3989 | arguments. |
3990 | |
3991 | The starting offset and size for this parm are returned in |
3992 | LOCATE->OFFSET and LOCATE->SIZE, respectively. When IN_REGS is |
3993 | nonzero, the offset is that of stack slot, which is returned in |
3994 | LOCATE->SLOT_OFFSET. LOCATE->ALIGNMENT_PAD is the amount of |
3995 | padding required from the initial offset ptr to the stack slot. |
3996 | |
3997 | IN_REGS is nonzero if the argument will be passed in registers. It will |
3998 | never be set if REG_PARM_STACK_SPACE is not defined. |
3999 | |
4000 | REG_PARM_STACK_SPACE is the number of bytes of stack space reserved |
4001 | for arguments which are passed in registers. |
4002 | |
4003 | FNDECL is the function in which the argument was defined. |
4004 | |
4005 | There are two types of rounding that are done. The first, controlled by |
4006 | TARGET_FUNCTION_ARG_BOUNDARY, forces the offset from the start of the |
4007 | argument list to be aligned to the specific boundary (in bits). This |
4008 | rounding affects the initial and starting offsets, but not the argument |
4009 | size. |
4010 | |
4011 | The second, controlled by TARGET_FUNCTION_ARG_PADDING and PARM_BOUNDARY, |
4012 | optionally rounds the size of the parm to PARM_BOUNDARY. The |
4013 | initial offset is not affected by this rounding, while the size always |
4014 | is and the starting offset may be. */ |
4015 | |
4016 | /* LOCATE->OFFSET will be negative for ARGS_GROW_DOWNWARD case; |
4017 | INITIAL_OFFSET_PTR is positive because locate_and_pad_parm's |
4018 | callers pass in the total size of args so far as |
4019 | INITIAL_OFFSET_PTR. LOCATE->SIZE is always positive. */ |
4020 | |
4021 | void |
4022 | locate_and_pad_parm (machine_mode passed_mode, tree type, int in_regs, |
4023 | int reg_parm_stack_space, int partial, |
4024 | tree fndecl ATTRIBUTE_UNUSED, |
4025 | struct args_size *initial_offset_ptr, |
4026 | struct locate_and_pad_arg_data *locate) |
4027 | { |
4028 | tree sizetree; |
4029 | pad_direction where_pad; |
4030 | unsigned int boundary, round_boundary; |
4031 | int part_size_in_regs; |
4032 | |
4033 | /* If we have found a stack parm before we reach the end of the |
4034 | area reserved for registers, skip that area. */ |
4035 | if (! in_regs) |
4036 | { |
4037 | if (reg_parm_stack_space > 0) |
4038 | { |
4039 | if (initial_offset_ptr->var |
4040 | || !ordered_p (a: initial_offset_ptr->constant, |
4041 | b: reg_parm_stack_space)) |
4042 | { |
4043 | initial_offset_ptr->var |
4044 | = size_binop (MAX_EXPR, ARGS_SIZE_TREE (*initial_offset_ptr), |
4045 | ssize_int (reg_parm_stack_space)); |
4046 | initial_offset_ptr->constant = 0; |
4047 | } |
4048 | else |
4049 | initial_offset_ptr->constant |
4050 | = ordered_max (a: initial_offset_ptr->constant, |
4051 | b: reg_parm_stack_space); |
4052 | } |
4053 | } |
4054 | |
4055 | part_size_in_regs = (reg_parm_stack_space == 0 ? partial : 0); |
4056 | |
4057 | sizetree = (type |
4058 | ? arg_size_in_bytes (type) |
4059 | : size_int (GET_MODE_SIZE (passed_mode))); |
4060 | where_pad = targetm.calls.function_arg_padding (passed_mode, type); |
4061 | boundary = targetm.calls.function_arg_boundary (passed_mode, type); |
4062 | round_boundary = targetm.calls.function_arg_round_boundary (passed_mode, |
4063 | type); |
4064 | locate->where_pad = where_pad; |
4065 | |
4066 | /* Alignment can't exceed MAX_SUPPORTED_STACK_ALIGNMENT. */ |
4067 | if (boundary > MAX_SUPPORTED_STACK_ALIGNMENT) |
4068 | boundary = MAX_SUPPORTED_STACK_ALIGNMENT; |
4069 | |
4070 | locate->boundary = boundary; |
4071 | |
4072 | if (SUPPORTS_STACK_ALIGNMENT) |
4073 | { |
4074 | /* stack_alignment_estimated can't change after stack has been |
4075 | realigned. */ |
4076 | if (crtl->stack_alignment_estimated < boundary) |
4077 | { |
4078 | if (!crtl->stack_realign_processed) |
4079 | crtl->stack_alignment_estimated = boundary; |
4080 | else |
4081 | { |
4082 | /* If stack is realigned and stack alignment value |
4083 | hasn't been finalized, it is OK not to increase |
4084 | stack_alignment_estimated. The bigger alignment |
4085 | requirement is recorded in stack_alignment_needed |
4086 | below. */ |
4087 | gcc_assert (!crtl->stack_realign_finalized |
4088 | && crtl->stack_realign_needed); |
4089 | } |
4090 | } |
4091 | } |
4092 | |
4093 | if (ARGS_GROW_DOWNWARD) |
4094 | { |
4095 | locate->slot_offset.constant = -initial_offset_ptr->constant; |
4096 | if (initial_offset_ptr->var) |
4097 | locate->slot_offset.var = size_binop (MINUS_EXPR, ssize_int (0), |
4098 | initial_offset_ptr->var); |
4099 | |
4100 | { |
4101 | tree s2 = sizetree; |
4102 | if (where_pad != PAD_NONE |
4103 | && (!tree_fits_uhwi_p (sizetree) |
4104 | || (tree_to_uhwi (sizetree) * BITS_PER_UNIT) % round_boundary)) |
4105 | s2 = round_up (s2, round_boundary / BITS_PER_UNIT); |
4106 | SUB_PARM_SIZE (locate->slot_offset, s2); |
4107 | } |
4108 | |
4109 | locate->slot_offset.constant += part_size_in_regs; |
4110 | |
4111 | if (!in_regs || reg_parm_stack_space > 0) |
4112 | pad_to_arg_alignment (&locate->slot_offset, boundary, |
4113 | &locate->alignment_pad); |
4114 | |
4115 | locate->size.constant = (-initial_offset_ptr->constant |
4116 | - locate->slot_offset.constant); |
4117 | if (initial_offset_ptr->var) |
4118 | locate->size.var = size_binop (MINUS_EXPR, |
4119 | size_binop (MINUS_EXPR, |
4120 | ssize_int (0), |
4121 | initial_offset_ptr->var), |
4122 | locate->slot_offset.var); |
4123 | |
4124 | /* Pad_below needs the pre-rounded size to know how much to pad |
4125 | below. */ |
4126 | locate->offset = locate->slot_offset; |
4127 | if (where_pad == PAD_DOWNWARD) |
4128 | pad_below (&locate->offset, passed_mode, sizetree); |
4129 | |
4130 | } |
4131 | else |
4132 | { |
4133 | if (!in_regs || reg_parm_stack_space > 0) |
4134 | pad_to_arg_alignment (initial_offset_ptr, boundary, |
4135 | &locate->alignment_pad); |
4136 | locate->slot_offset = *initial_offset_ptr; |
4137 | |
4138 | #ifdef PUSH_ROUNDING |
4139 | if (passed_mode != BLKmode) |
4140 | sizetree = size_int (PUSH_ROUNDING (TREE_INT_CST_LOW (sizetree))); |
4141 | #endif |
4142 | |
4143 | /* Pad_below needs the pre-rounded size to know how much to pad below |
4144 | so this must be done before rounding up. */ |
4145 | locate->offset = locate->slot_offset; |
4146 | if (where_pad == PAD_DOWNWARD) |
4147 | pad_below (&locate->offset, passed_mode, sizetree); |
4148 | |
4149 | if (where_pad != PAD_NONE |
4150 | && (!tree_fits_uhwi_p (sizetree) |
4151 | || (tree_to_uhwi (sizetree) * BITS_PER_UNIT) % round_boundary)) |
4152 | sizetree = round_up (sizetree, round_boundary / BITS_PER_UNIT); |
4153 | |
4154 | ADD_PARM_SIZE (locate->size, sizetree); |
4155 | |
4156 | locate->size.constant -= part_size_in_regs; |
4157 | } |
4158 | |
4159 | locate->offset.constant |
4160 | += targetm.calls.function_arg_offset (passed_mode, type); |
4161 | } |
4162 | |
4163 | /* Round the stack offset in *OFFSET_PTR up to a multiple of BOUNDARY. |
4164 | BOUNDARY is measured in bits, but must be a multiple of a storage unit. */ |
4165 | |
4166 | static void |
4167 | pad_to_arg_alignment (struct args_size *offset_ptr, int boundary, |
4168 | struct args_size *alignment_pad) |
4169 | { |
4170 | tree save_var = NULL_TREE; |
4171 | poly_int64 save_constant = 0; |
4172 | int boundary_in_bytes = boundary / BITS_PER_UNIT; |
4173 | poly_int64 sp_offset = STACK_POINTER_OFFSET; |
4174 | |
4175 | #ifdef SPARC_STACK_BOUNDARY_HACK |
4176 | /* ??? The SPARC port may claim a STACK_BOUNDARY higher than |
4177 | the real alignment of %sp. However, when it does this, the |
4178 | alignment of %sp+STACK_POINTER_OFFSET is STACK_BOUNDARY. */ |
4179 | if (SPARC_STACK_BOUNDARY_HACK) |
4180 | sp_offset = 0; |
4181 | #endif |
4182 | |
4183 | if (boundary > PARM_BOUNDARY) |
4184 | { |
4185 | save_var = offset_ptr->var; |
4186 | save_constant = offset_ptr->constant; |
4187 | } |
4188 | |
4189 | alignment_pad->var = NULL_TREE; |
4190 | alignment_pad->constant = 0; |
4191 | |
4192 | if (boundary > BITS_PER_UNIT) |
4193 | { |
4194 | int misalign; |
4195 | if (offset_ptr->var |
4196 | || !known_misalignment (value: offset_ptr->constant + sp_offset, |
4197 | align: boundary_in_bytes, misalign: &misalign)) |
4198 | { |
4199 | tree sp_offset_tree = ssize_int (sp_offset); |
4200 | tree offset = size_binop (PLUS_EXPR, |
4201 | ARGS_SIZE_TREE (*offset_ptr), |
4202 | sp_offset_tree); |
4203 | tree rounded; |
4204 | if (ARGS_GROW_DOWNWARD) |
4205 | rounded = round_down (offset, boundary / BITS_PER_UNIT); |
4206 | else |
4207 | rounded = round_up (offset, boundary / BITS_PER_UNIT); |
4208 | |
4209 | offset_ptr->var = size_binop (MINUS_EXPR, rounded, sp_offset_tree); |
4210 | /* ARGS_SIZE_TREE includes constant term. */ |
4211 | offset_ptr->constant = 0; |
4212 | if (boundary > PARM_BOUNDARY) |
4213 | alignment_pad->var = size_binop (MINUS_EXPR, offset_ptr->var, |
4214 | save_var); |
4215 | } |
4216 | else |
4217 | { |
4218 | if (ARGS_GROW_DOWNWARD) |
4219 | offset_ptr->constant -= misalign; |
4220 | else |
4221 | offset_ptr->constant += -misalign & (boundary_in_bytes - 1); |
4222 | |
4223 | if (boundary > PARM_BOUNDARY) |
4224 | alignment_pad->constant = offset_ptr->constant - save_constant; |
4225 | } |
4226 | } |
4227 | } |
4228 | |
4229 | static void |
4230 | pad_below (struct args_size *offset_ptr, machine_mode passed_mode, tree sizetree) |
4231 | { |
4232 | unsigned int align = PARM_BOUNDARY / BITS_PER_UNIT; |
4233 | int misalign; |
4234 | if (passed_mode != BLKmode |
4235 | && known_misalignment (value: GET_MODE_SIZE (mode: passed_mode), align, misalign: &misalign)) |
4236 | offset_ptr->constant += -misalign & (align - 1); |
4237 | else |
4238 | { |
4239 | if (TREE_CODE (sizetree) != INTEGER_CST |
4240 | || (TREE_INT_CST_LOW (sizetree) & (align - 1)) != 0) |
4241 | { |
4242 | /* Round the size up to multiple of PARM_BOUNDARY bits. */ |
4243 | tree s2 = round_up (sizetree, align); |
4244 | /* Add it in. */ |
4245 | ADD_PARM_SIZE (*offset_ptr, s2); |
4246 | SUB_PARM_SIZE (*offset_ptr, sizetree); |
4247 | } |
4248 | } |
4249 | } |
4250 | |
4251 | |
4252 | /* True if register REGNO was alive at a place where `setjmp' was |
4253 | called and was set more than once or is an argument. Such regs may |
4254 | be clobbered by `longjmp'. */ |
4255 | |
4256 | static bool |
4257 | regno_clobbered_at_setjmp (bitmap setjmp_crosses, int regno) |
4258 | { |
4259 | /* There appear to be cases where some local vars never reach the |
4260 | backend but have bogus regnos. */ |
4261 | if (regno >= max_reg_num ()) |
4262 | return false; |
4263 | |
4264 | return ((REG_N_SETS (regno) > 1 |
4265 | || REGNO_REG_SET_P (df_get_live_out (ENTRY_BLOCK_PTR_FOR_FN (cfun)), |
4266 | regno)) |
4267 | && REGNO_REG_SET_P (setjmp_crosses, regno)); |
4268 | } |
4269 | |
4270 | /* Walk the tree of blocks describing the binding levels within a |
4271 | function and warn about variables the might be killed by setjmp or |
4272 | vfork. This is done after calling flow_analysis before register |
4273 | allocation since that will clobber the pseudo-regs to hard |
4274 | regs. */ |
4275 | |
4276 | static void |
4277 | setjmp_vars_warning (bitmap setjmp_crosses, tree block) |
4278 | { |
4279 | tree decl, sub; |
4280 | |
4281 | for (decl = BLOCK_VARS (block); decl; decl = DECL_CHAIN (decl)) |
4282 | { |
4283 | if (VAR_P (decl) |
4284 | && DECL_RTL_SET_P (decl) |
4285 | && REG_P (DECL_RTL (decl)) |
4286 | && regno_clobbered_at_setjmp (setjmp_crosses, REGNO (DECL_RTL (decl)))) |
4287 | warning (OPT_Wclobbered, "variable %q+D might be clobbered by" |
4288 | " %<longjmp%> or %<vfork%>" , decl); |
4289 | } |
4290 | |
4291 | for (sub = BLOCK_SUBBLOCKS (block); sub; sub = BLOCK_CHAIN (sub)) |
4292 | setjmp_vars_warning (setjmp_crosses, block: sub); |
4293 | } |
4294 | |
4295 | /* Do the appropriate part of setjmp_vars_warning |
4296 | but for arguments instead of local variables. */ |
4297 | |
4298 | static void |
4299 | setjmp_args_warning (bitmap setjmp_crosses) |
4300 | { |
4301 | tree decl; |
4302 | for (decl = DECL_ARGUMENTS (current_function_decl); |
4303 | decl; decl = DECL_CHAIN (decl)) |
4304 | if (DECL_RTL (decl) != 0 |
4305 | && REG_P (DECL_RTL (decl)) |
4306 | && regno_clobbered_at_setjmp (setjmp_crosses, REGNO (DECL_RTL (decl)))) |
4307 | warning (OPT_Wclobbered, |
4308 | "argument %q+D might be clobbered by %<longjmp%> or %<vfork%>" , |
4309 | decl); |
4310 | } |
4311 | |
4312 | /* Generate warning messages for variables live across setjmp. */ |
4313 | |
4314 | void |
4315 | generate_setjmp_warnings (void) |
4316 | { |
4317 | bitmap setjmp_crosses = regstat_get_setjmp_crosses (); |
4318 | |
4319 | if (n_basic_blocks_for_fn (cfun) == NUM_FIXED_BLOCKS |
4320 | || bitmap_empty_p (map: setjmp_crosses)) |
4321 | return; |
4322 | |
4323 | setjmp_vars_warning (setjmp_crosses, DECL_INITIAL (current_function_decl)); |
4324 | setjmp_args_warning (setjmp_crosses); |
4325 | } |
4326 | |
4327 | |
4328 | /* Reverse the order of elements in the fragment chain T of blocks, |
4329 | and return the new head of the chain (old last element). |
4330 | In addition to that clear BLOCK_SAME_RANGE flags when needed |
4331 | and adjust BLOCK_SUPERCONTEXT from the super fragment to |
4332 | its super fragment origin. */ |
4333 | |
4334 | static tree |
4335 | block_fragments_nreverse (tree t) |
4336 | { |
4337 | tree prev = 0, block, next, prev_super = 0; |
4338 | tree super = BLOCK_SUPERCONTEXT (t); |
4339 | if (BLOCK_FRAGMENT_ORIGIN (super)) |
4340 | super = BLOCK_FRAGMENT_ORIGIN (super); |
4341 | for (block = t; block; block = next) |
4342 | { |
4343 | next = BLOCK_FRAGMENT_CHAIN (block); |
4344 | BLOCK_FRAGMENT_CHAIN (block) = prev; |
4345 | if ((prev && !BLOCK_SAME_RANGE (prev)) |
4346 | || (BLOCK_FRAGMENT_CHAIN (BLOCK_SUPERCONTEXT (block)) |
4347 | != prev_super)) |
4348 | BLOCK_SAME_RANGE (block) = 0; |
4349 | prev_super = BLOCK_SUPERCONTEXT (block); |
4350 | BLOCK_SUPERCONTEXT (block) = super; |
4351 | prev = block; |
4352 | } |
4353 | t = BLOCK_FRAGMENT_ORIGIN (t); |
4354 | if (BLOCK_FRAGMENT_CHAIN (BLOCK_SUPERCONTEXT (t)) |
4355 | != prev_super) |
4356 | BLOCK_SAME_RANGE (t) = 0; |
4357 | BLOCK_SUPERCONTEXT (t) = super; |
4358 | return prev; |
4359 | } |
4360 | |
4361 | /* Reverse the order of elements in the chain T of blocks, |
4362 | and return the new head of the chain (old last element). |
4363 | Also do the same on subblocks and reverse the order of elements |
4364 | in BLOCK_FRAGMENT_CHAIN as well. */ |
4365 | |
4366 | static tree |
4367 | blocks_nreverse_all (tree t) |
4368 | { |
4369 | tree prev = 0, block, next; |
4370 | for (block = t; block; block = next) |
4371 | { |
4372 | next = BLOCK_CHAIN (block); |
4373 | BLOCK_CHAIN (block) = prev; |
4374 | if (BLOCK_FRAGMENT_CHAIN (block) |
4375 | && BLOCK_FRAGMENT_ORIGIN (block) == NULL_TREE) |
4376 | { |
4377 | BLOCK_FRAGMENT_CHAIN (block) |
4378 | = block_fragments_nreverse (BLOCK_FRAGMENT_CHAIN (block)); |
4379 | if (!BLOCK_SAME_RANGE (BLOCK_FRAGMENT_CHAIN (block))) |
4380 | BLOCK_SAME_RANGE (block) = 0; |
4381 | } |
4382 | BLOCK_SUBBLOCKS (block) = blocks_nreverse_all (BLOCK_SUBBLOCKS (block)); |
4383 | prev = block; |
4384 | } |
4385 | return prev; |
4386 | } |
4387 | |
4388 | |
4389 | /* Identify BLOCKs referenced by more than one NOTE_INSN_BLOCK_{BEG,END}, |
4390 | and create duplicate blocks. */ |
4391 | /* ??? Need an option to either create block fragments or to create |
4392 | abstract origin duplicates of a source block. It really depends |
4393 | on what optimization has been performed. */ |
4394 | |
4395 | void |
4396 | reorder_blocks (void) |
4397 | { |
4398 | tree block = DECL_INITIAL (current_function_decl); |
4399 | |
4400 | if (block == NULL_TREE) |
4401 | return; |
4402 | |
4403 | auto_vec<tree, 10> block_stack; |
4404 | |
4405 | /* Reset the TREE_ASM_WRITTEN bit for all blocks. */ |
4406 | clear_block_marks (block); |
4407 | |
4408 | /* Prune the old trees away, so that they don't get in the way. */ |
4409 | BLOCK_SUBBLOCKS (block) = NULL_TREE; |
4410 | BLOCK_CHAIN (block) = NULL_TREE; |
4411 | |
4412 | /* Recreate the block tree from the note nesting. */ |
4413 | reorder_blocks_1 (get_insns (), block, &block_stack); |
4414 | BLOCK_SUBBLOCKS (block) = blocks_nreverse_all (BLOCK_SUBBLOCKS (block)); |
4415 | } |
4416 | |
4417 | /* Helper function for reorder_blocks. Reset TREE_ASM_WRITTEN. */ |
4418 | |
4419 | void |
4420 | clear_block_marks (tree block) |
4421 | { |
4422 | while (block) |
4423 | { |
4424 | TREE_ASM_WRITTEN (block) = 0; |
4425 | clear_block_marks (BLOCK_SUBBLOCKS (block)); |
4426 | block = BLOCK_CHAIN (block); |
4427 | } |
4428 | } |
4429 | |
4430 | static void |
4431 | reorder_blocks_1 (rtx_insn *insns, tree current_block, |
4432 | vec<tree> *p_block_stack) |
4433 | { |
4434 | rtx_insn *insn; |
4435 | tree prev_beg = NULL_TREE, prev_end = NULL_TREE; |
4436 | |
4437 | for (insn = insns; insn; insn = NEXT_INSN (insn)) |
4438 | { |
4439 | if (NOTE_P (insn)) |
4440 | { |
4441 | if (NOTE_KIND (insn) == NOTE_INSN_BLOCK_BEG) |
4442 | { |
4443 | tree block = NOTE_BLOCK (insn); |
4444 | tree origin; |
4445 | |
4446 | gcc_assert (BLOCK_FRAGMENT_ORIGIN (block) == NULL_TREE); |
4447 | origin = block; |
4448 | |
4449 | if (prev_end) |
4450 | BLOCK_SAME_RANGE (prev_end) = 0; |
4451 | prev_end = NULL_TREE; |
4452 | |
4453 | /* If we have seen this block before, that means it now |
4454 | spans multiple address regions. Create a new fragment. */ |
4455 | if (TREE_ASM_WRITTEN (block)) |
4456 | { |
4457 | tree new_block = copy_node (block); |
4458 | |
4459 | BLOCK_SAME_RANGE (new_block) = 0; |
4460 | BLOCK_FRAGMENT_ORIGIN (new_block) = origin; |
4461 | BLOCK_FRAGMENT_CHAIN (new_block) |
4462 | = BLOCK_FRAGMENT_CHAIN (origin); |
4463 | BLOCK_FRAGMENT_CHAIN (origin) = new_block; |
4464 | |
4465 | NOTE_BLOCK (insn) = new_block; |
4466 | block = new_block; |
4467 | } |
4468 | |
4469 | if (prev_beg == current_block && prev_beg) |
4470 | BLOCK_SAME_RANGE (block) = 1; |
4471 | |
4472 | prev_beg = origin; |
4473 | |
4474 | BLOCK_SUBBLOCKS (block) = 0; |
4475 | TREE_ASM_WRITTEN (block) = 1; |
4476 | /* When there's only one block for the entire function, |
4477 | current_block == block and we mustn't do this, it |
4478 | will cause infinite recursion. */ |
4479 | if (block != current_block) |
4480 | { |
4481 | tree super; |
4482 | if (block != origin) |
4483 | gcc_assert (BLOCK_SUPERCONTEXT (origin) == current_block |
4484 | || BLOCK_FRAGMENT_ORIGIN (BLOCK_SUPERCONTEXT |
4485 | (origin)) |
4486 | == current_block); |
4487 | if (p_block_stack->is_empty ()) |
4488 | super = current_block; |
4489 | else |
4490 | { |
4491 | super = p_block_stack->last (); |
4492 | gcc_assert (super == current_block |
4493 | || BLOCK_FRAGMENT_ORIGIN (super) |
4494 | == current_block); |
4495 | } |
4496 | BLOCK_SUPERCONTEXT (block) = super; |
4497 | BLOCK_CHAIN (block) = BLOCK_SUBBLOCKS (current_block); |
4498 | BLOCK_SUBBLOCKS (current_block) = block; |
4499 | current_block = origin; |
4500 | } |
4501 | p_block_stack->safe_push (obj: block); |
4502 | } |
4503 | else if (NOTE_KIND (insn) == NOTE_INSN_BLOCK_END) |
4504 | { |
4505 | NOTE_BLOCK (insn) = p_block_stack->pop (); |
4506 | current_block = BLOCK_SUPERCONTEXT (current_block); |
4507 | if (BLOCK_FRAGMENT_ORIGIN (current_block)) |
4508 | current_block = BLOCK_FRAGMENT_ORIGIN (current_block); |
4509 | prev_beg = NULL_TREE; |
4510 | prev_end = BLOCK_SAME_RANGE (NOTE_BLOCK (insn)) |
4511 | ? NOTE_BLOCK (insn) : NULL_TREE; |
4512 | } |
4513 | } |
4514 | else |
4515 | { |
4516 | prev_beg = NULL_TREE; |
4517 | if (prev_end) |
4518 | BLOCK_SAME_RANGE (prev_end) = 0; |
4519 | prev_end = NULL_TREE; |
4520 | } |
4521 | } |
4522 | } |
4523 | |
4524 | /* Reverse the order of elements in the chain T of blocks, |
4525 | and return the new head of the chain (old last element). */ |
4526 | |
4527 | tree |
4528 | blocks_nreverse (tree t) |
4529 | { |
4530 | tree prev = 0, block, next; |
4531 | for (block = t; block; block = next) |
4532 | { |
4533 | next = BLOCK_CHAIN (block); |
4534 | BLOCK_CHAIN (block) = prev; |
4535 | prev = block; |
4536 | } |
4537 | return prev; |
4538 | } |
4539 | |
4540 | /* Concatenate two chains of blocks (chained through BLOCK_CHAIN) |
4541 | by modifying the last node in chain 1 to point to chain 2. */ |
4542 | |
4543 | tree |
4544 | block_chainon (tree op1, tree op2) |
4545 | { |
4546 | tree t1; |
4547 | |
4548 | if (!op1) |
4549 | return op2; |
4550 | if (!op2) |
4551 | return op1; |
4552 | |
4553 | for (t1 = op1; BLOCK_CHAIN (t1); t1 = BLOCK_CHAIN (t1)) |
4554 | continue; |
4555 | BLOCK_CHAIN (t1) = op2; |
4556 | |
4557 | #ifdef ENABLE_TREE_CHECKING |
4558 | { |
4559 | tree t2; |
4560 | for (t2 = op2; t2; t2 = BLOCK_CHAIN (t2)) |
4561 | gcc_assert (t2 != t1); |
4562 | } |
4563 | #endif |
4564 | |
4565 | return op1; |
4566 | } |
4567 | |
4568 | /* Count the subblocks of the list starting with BLOCK. If VECTOR is |
4569 | non-NULL, list them all into VECTOR, in a depth-first preorder |
4570 | traversal of the block tree. Also clear TREE_ASM_WRITTEN in all |
4571 | blocks. */ |
4572 | |
4573 | static int |
4574 | all_blocks (tree block, tree *vector) |
4575 | { |
4576 | int n_blocks = 0; |
4577 | |
4578 | while (block) |
4579 | { |
4580 | TREE_ASM_WRITTEN (block) = 0; |
4581 | |
4582 | /* Record this block. */ |
4583 | if (vector) |
4584 | vector[n_blocks] = block; |
4585 | |
4586 | ++n_blocks; |
4587 | |
4588 | /* Record the subblocks, and their subblocks... */ |
4589 | n_blocks += all_blocks (BLOCK_SUBBLOCKS (block), |
4590 | vector: vector ? vector + n_blocks : 0); |
4591 | block = BLOCK_CHAIN (block); |
4592 | } |
4593 | |
4594 | return n_blocks; |
4595 | } |
4596 | |
4597 | /* Return a vector containing all the blocks rooted at BLOCK. The |
4598 | number of elements in the vector is stored in N_BLOCKS_P. The |
4599 | vector is dynamically allocated; it is the caller's responsibility |
4600 | to call `free' on the pointer returned. */ |
4601 | |
4602 | static tree * |
4603 | get_block_vector (tree block, int *n_blocks_p) |
4604 | { |
4605 | tree *block_vector; |
4606 | |
4607 | *n_blocks_p = all_blocks (block, NULL); |
4608 | block_vector = XNEWVEC (tree, *n_blocks_p); |
4609 | all_blocks (block, vector: block_vector); |
4610 | |
4611 | return block_vector; |
4612 | } |
4613 | |
4614 | static GTY(()) int next_block_index = 2; |
4615 | |
4616 | /* Set BLOCK_NUMBER for all the blocks in FN. */ |
4617 | |
4618 | void |
4619 | number_blocks (tree fn) |
4620 | { |
4621 | int i; |
4622 | int n_blocks; |
4623 | tree *block_vector; |
4624 | |
4625 | block_vector = get_block_vector (DECL_INITIAL (fn), n_blocks_p: &n_blocks); |
4626 | |
4627 | /* The top-level BLOCK isn't numbered at all. */ |
4628 | for (i = 1; i < n_blocks; ++i) |
4629 | /* We number the blocks from two. */ |
4630 | BLOCK_NUMBER (block_vector[i]) = next_block_index++; |
4631 | |
4632 | free (ptr: block_vector); |
4633 | |
4634 | return; |
4635 | } |
4636 | |
4637 | /* If VAR is present in a subblock of BLOCK, return the subblock. */ |
4638 | |
4639 | DEBUG_FUNCTION tree |
4640 | debug_find_var_in_block_tree (tree var, tree block) |
4641 | { |
4642 | tree t; |
4643 | |
4644 | for (t = BLOCK_VARS (block); t; t = TREE_CHAIN (t)) |
4645 | if (t == var) |
4646 | return block; |
4647 | |
4648 | for (t = BLOCK_SUBBLOCKS (block); t; t = TREE_CHAIN (t)) |
4649 | { |
4650 | tree ret = debug_find_var_in_block_tree (var, block: t); |
4651 | if (ret) |
4652 | return ret; |
4653 | } |
4654 | |
4655 | return NULL_TREE; |
4656 | } |
4657 | |
4658 | /* Keep track of whether we're in a dummy function context. If we are, |
4659 | we don't want to invoke the set_current_function hook, because we'll |
4660 | get into trouble if the hook calls target_reinit () recursively or |
4661 | when the initial initialization is not yet complete. */ |
4662 | |
4663 | static bool in_dummy_function; |
4664 | |
4665 | /* Invoke the target hook when setting cfun. Update the optimization options |
4666 | if the function uses different options than the default. */ |
4667 | |
4668 | static void |
4669 | invoke_set_current_function_hook (tree fndecl) |
4670 | { |
4671 | if (!in_dummy_function) |
4672 | { |
4673 | tree opts = ((fndecl) |
4674 | ? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl) |
4675 | : optimization_default_node); |
4676 | |
4677 | if (!opts) |
4678 | opts = optimization_default_node; |
4679 | |
4680 | /* Change optimization options if needed. */ |
4681 | if (optimization_current_node != opts) |
4682 | { |
4683 | optimization_current_node = opts; |
4684 | cl_optimization_restore (&global_options, &global_options_set, |
4685 | TREE_OPTIMIZATION (opts)); |
4686 | } |
4687 | |
4688 | targetm.set_current_function (fndecl); |
4689 | this_fn_optabs = this_target_optabs; |
4690 | |
4691 | /* Initialize global alignment variables after op. */ |
4692 | parse_alignment_opts (); |
4693 | |
4694 | if (opts != optimization_default_node) |
4695 | { |
4696 | init_tree_optimization_optabs (opts); |
4697 | if (TREE_OPTIMIZATION_OPTABS (opts)) |
4698 | this_fn_optabs = (struct target_optabs *) |
4699 | TREE_OPTIMIZATION_OPTABS (opts); |
4700 | } |
4701 | } |
4702 | } |
4703 | |
4704 | /* cfun should never be set directly; use this function. */ |
4705 | |
4706 | void |
4707 | set_cfun (struct function *new_cfun, bool force) |
4708 | { |
4709 | if (cfun != new_cfun || force) |
4710 | { |
4711 | cfun = new_cfun; |
4712 | invoke_set_current_function_hook (fndecl: new_cfun ? new_cfun->decl : NULL_TREE); |
4713 | redirect_edge_var_map_empty (); |
4714 | } |
4715 | } |
4716 | |
4717 | /* Initialized with NOGC, making this poisonous to the garbage collector. */ |
4718 | |
4719 | static vec<function *> cfun_stack; |
4720 | |
4721 | /* Push the current cfun onto the stack, and set cfun to new_cfun. Also set |
4722 | current_function_decl accordingly. */ |
4723 | |
4724 | void |
4725 | push_cfun (struct function *new_cfun) |
4726 | { |
4727 | gcc_assert ((!cfun && !current_function_decl) |
4728 | || (cfun && current_function_decl == cfun->decl)); |
4729 | cfun_stack.safe_push (obj: cfun); |
4730 | current_function_decl = new_cfun ? new_cfun->decl : NULL_TREE; |
4731 | set_cfun (new_cfun); |
4732 | } |
4733 | |
4734 | /* Pop cfun from the stack. Also set current_function_decl accordingly. */ |
4735 | |
4736 | void |
4737 | pop_cfun (void) |
4738 | { |
4739 | struct function *new_cfun = cfun_stack.pop (); |
4740 | /* When in_dummy_function, we do have a cfun but current_function_decl is |
4741 | NULL. We also allow pushing NULL cfun and subsequently changing |
4742 | current_function_decl to something else and have both restored by |
4743 | pop_cfun. */ |
4744 | gcc_checking_assert (in_dummy_function |
4745 | || !cfun |
4746 | || current_function_decl == cfun->decl); |
4747 | set_cfun (new_cfun); |
4748 | current_function_decl = new_cfun ? new_cfun->decl : NULL_TREE; |
4749 | } |
4750 | |
4751 | /* Return value of funcdef and increase it. */ |
4752 | int |
4753 | get_next_funcdef_no (void) |
4754 | { |
4755 | return funcdef_no++; |
4756 | } |
4757 | |
4758 | /* Return value of funcdef. */ |
4759 | int |
4760 | get_last_funcdef_no (void) |
4761 | { |
4762 | return funcdef_no; |
4763 | } |
4764 | |
4765 | /* Allocate and initialize the stack usage info data structure for the |
4766 | current function. */ |
4767 | static void |
4768 | allocate_stack_usage_info (void) |
4769 | { |
4770 | gcc_assert (!cfun->su); |
4771 | cfun->su = ggc_cleared_alloc<stack_usage> (); |
4772 | cfun->su->static_stack_size = -1; |
4773 | } |
4774 | |
4775 | /* Allocate a function structure for FNDECL and set its contents |
4776 | to the defaults. Set cfun to the newly-allocated object. |
4777 | Some of the helper functions invoked during initialization assume |
4778 | that cfun has already been set. Therefore, assign the new object |
4779 | directly into cfun and invoke the back end hook explicitly at the |
4780 | very end, rather than initializing a temporary and calling set_cfun |
4781 | on it. |
4782 | |
4783 | ABSTRACT_P is true if this is a function that will never be seen by |
4784 | the middle-end. Such functions are front-end concepts (like C++ |
4785 | function templates) that do not correspond directly to functions |
4786 | placed in object files. */ |
4787 | |
4788 | void |
4789 | allocate_struct_function (tree fndecl, bool abstract_p) |
4790 | { |
4791 | tree fntype = fndecl ? TREE_TYPE (fndecl) : NULL_TREE; |
4792 | |
4793 | cfun = ggc_cleared_alloc<function> (); |
4794 | |
4795 | init_eh_for_function (); |
4796 | |
4797 | if (init_machine_status) |
4798 | cfun->machine = (*init_machine_status) (); |
4799 | |
4800 | #ifdef OVERRIDE_ABI_FORMAT |
4801 | OVERRIDE_ABI_FORMAT (fndecl); |
4802 | #endif |
4803 | |
4804 | if (fndecl != NULL_TREE) |
4805 | { |
4806 | DECL_STRUCT_FUNCTION (fndecl) = cfun; |
4807 | cfun->decl = fndecl; |
4808 | current_function_funcdef_no = get_next_funcdef_no (); |
4809 | } |
4810 | |
4811 | invoke_set_current_function_hook (fndecl); |
4812 | |
4813 | if (fndecl != NULL_TREE) |
4814 | { |
4815 | tree result = DECL_RESULT (fndecl); |
4816 | |
4817 | if (!abstract_p) |
4818 | { |
4819 | /* Now that we have activated any function-specific attributes |
4820 | that might affect layout, particularly vector modes, relayout |
4821 | each of the parameters and the result. */ |
4822 | relayout_decl (result); |
4823 | for (tree parm = DECL_ARGUMENTS (fndecl); parm; |
4824 | parm = DECL_CHAIN (parm)) |
4825 | relayout_decl (parm); |
4826 | |
4827 | /* Similarly relayout the function decl. */ |
4828 | targetm.target_option.relayout_function (fndecl); |
4829 | } |
4830 | |
4831 | if (!abstract_p && aggregate_value_p (exp: result, fntype: fndecl)) |
4832 | { |
4833 | #ifdef PCC_STATIC_STRUCT_RETURN |
4834 | cfun->returns_pcc_struct = 1; |
4835 | #endif |
4836 | cfun->returns_struct = 1; |
4837 | } |
4838 | |
4839 | cfun->stdarg = stdarg_p (fntype); |
4840 | |
4841 | /* Assume all registers in stdarg functions need to be saved. */ |
4842 | cfun->va_list_gpr_size = VA_LIST_MAX_GPR_SIZE; |
4843 | cfun->va_list_fpr_size = VA_LIST_MAX_FPR_SIZE; |
4844 | |
4845 | /* ??? This could be set on a per-function basis by the front-end |
4846 | but is this worth the hassle? */ |
4847 | cfun->can_throw_non_call_exceptions = flag_non_call_exceptions; |
4848 | cfun->can_delete_dead_exceptions = flag_delete_dead_exceptions; |
4849 | |
4850 | if (!profile_flag && !flag_instrument_function_entry_exit) |
4851 | DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (fndecl) = 1; |
4852 | |
4853 | if (flag_callgraph_info) |
4854 | allocate_stack_usage_info (); |
4855 | } |
4856 | |
4857 | /* Don't enable begin stmt markers if var-tracking at assignments is |
4858 | disabled. The markers make little sense without the variable |
4859 | binding annotations among them. */ |
4860 | cfun->debug_nonbind_markers = lang_hooks.emits_begin_stmt |
4861 | && MAY_HAVE_DEBUG_MARKER_STMTS; |
4862 | } |
4863 | |
4864 | /* This is like allocate_struct_function, but pushes a new cfun for FNDECL |
4865 | instead of just setting it. */ |
4866 | |
4867 | void |
4868 | push_struct_function (tree fndecl, bool abstract_p) |
4869 | { |
4870 | /* When in_dummy_function we might be in the middle of a pop_cfun and |
4871 | current_function_decl and cfun may not match. */ |
4872 | gcc_assert (in_dummy_function |
4873 | || (!cfun && !current_function_decl) |
4874 | || (cfun && current_function_decl == cfun->decl)); |
4875 | cfun_stack.safe_push (obj: cfun); |
4876 | current_function_decl = fndecl; |
4877 | allocate_struct_function (fndecl, abstract_p); |
4878 | } |
4879 | |
4880 | /* Reset crtl and other non-struct-function variables to defaults as |
4881 | appropriate for emitting rtl at the start of a function. */ |
4882 | |
4883 | static void |
4884 | prepare_function_start (void) |
4885 | { |
4886 | gcc_assert (!get_last_insn ()); |
4887 | |
4888 | if (in_dummy_function) |
4889 | crtl->abi = &default_function_abi; |
4890 | else |
4891 | crtl->abi = &fndecl_abi (cfun->decl).base_abi (); |
4892 | |
4893 | init_temp_slots (); |
4894 | init_emit (); |
4895 | init_varasm_status (); |
4896 | init_expr (); |
4897 | default_rtl_profile (); |
4898 | |
4899 | if (flag_stack_usage_info && !flag_callgraph_info) |
4900 | allocate_stack_usage_info (); |
4901 | |
4902 | cse_not_expected = ! optimize; |
4903 | |
4904 | /* Caller save not needed yet. */ |
4905 | caller_save_needed = 0; |
4906 | |
4907 | /* We haven't done register allocation yet. */ |
4908 | reg_renumber = 0; |
4909 | |
4910 | /* Indicate that we have not instantiated virtual registers yet. */ |
4911 | virtuals_instantiated = 0; |
4912 | |
4913 | /* Indicate that we want CONCATs now. */ |
4914 | generating_concat_p = 1; |
4915 | |
4916 | /* Indicate we have no need of a frame pointer yet. */ |
4917 | frame_pointer_needed = 0; |
4918 | } |
4919 | |
4920 | void |
4921 | push_dummy_function (bool with_decl) |
4922 | { |
4923 | tree fn_decl, fn_type, fn_result_decl; |
4924 | |
4925 | gcc_assert (!in_dummy_function); |
4926 | in_dummy_function = true; |
4927 | |
4928 | if (with_decl) |
4929 | { |
4930 | fn_type = build_function_type_list (void_type_node, NULL_TREE); |
4931 | fn_decl = build_decl (UNKNOWN_LOCATION, FUNCTION_DECL, NULL_TREE, |
4932 | fn_type); |
4933 | fn_result_decl = build_decl (UNKNOWN_LOCATION, RESULT_DECL, |
4934 | NULL_TREE, void_type_node); |
4935 | DECL_RESULT (fn_decl) = fn_result_decl; |
4936 | DECL_ARTIFICIAL (fn_decl) = 1; |
4937 | tree fn_name = get_identifier (" " ); |
4938 | SET_DECL_ASSEMBLER_NAME (fn_decl, fn_name); |
4939 | } |
4940 | else |
4941 | fn_decl = NULL_TREE; |
4942 | |
4943 | push_struct_function (fndecl: fn_decl); |
4944 | } |
4945 | |
4946 | /* Initialize the rtl expansion mechanism so that we can do simple things |
4947 | like generate sequences. This is used to provide a context during global |
4948 | initialization of some passes. You must call expand_dummy_function_end |
4949 | to exit this context. */ |
4950 | |
4951 | void |
4952 | init_dummy_function_start (void) |
4953 | { |
4954 | push_dummy_function (with_decl: false); |
4955 | prepare_function_start (); |
4956 | } |
4957 | |
4958 | /* Generate RTL for the start of the function SUBR (a FUNCTION_DECL tree node) |
4959 | and initialize static variables for generating RTL for the statements |
4960 | of the function. */ |
4961 | |
4962 | void |
4963 | init_function_start (tree subr) |
4964 | { |
4965 | /* Initialize backend, if needed. */ |
4966 | initialize_rtl (); |
4967 | |
4968 | prepare_function_start (); |
4969 | decide_function_section (subr); |
4970 | |
4971 | /* Warn if this value is an aggregate type, |
4972 | regardless of which calling convention we are using for it. */ |
4973 | if (AGGREGATE_TYPE_P (TREE_TYPE (DECL_RESULT (subr)))) |
4974 | warning_at (DECL_SOURCE_LOCATION (DECL_RESULT (subr)), |
4975 | OPT_Waggregate_return, "function returns an aggregate" ); |
4976 | } |
4977 | |
4978 | /* Expand code to verify the stack_protect_guard. This is invoked at |
4979 | the end of a function to be protected. */ |
4980 | |
4981 | void |
4982 | stack_protect_epilogue (void) |
4983 | { |
4984 | tree guard_decl = crtl->stack_protect_guard_decl; |
4985 | rtx_code_label *label = gen_label_rtx (); |
4986 | rtx x, y; |
4987 | rtx_insn *seq = NULL; |
4988 | |
4989 | x = expand_normal (crtl->stack_protect_guard); |
4990 | |
4991 | if (targetm.have_stack_protect_combined_test () && guard_decl) |
4992 | { |
4993 | gcc_assert (DECL_P (guard_decl)); |
4994 | y = DECL_RTL (guard_decl); |
4995 | /* Allow the target to compute address of Y and compare it with X without |
4996 | leaking Y into a register. This combined address + compare pattern |
4997 | allows the target to prevent spilling of any intermediate results by |
4998 | splitting it after register allocator. */ |
4999 | seq = targetm.gen_stack_protect_combined_test (x, y, label); |
5000 | } |
5001 | else |
5002 | { |
5003 | if (guard_decl) |
5004 | y = expand_normal (exp: guard_decl); |
5005 | else |
5006 | y = const0_rtx; |
5007 | |
5008 | /* Allow the target to compare Y with X without leaking either into |
5009 | a register. */ |
5010 | if (targetm.have_stack_protect_test ()) |
5011 | seq = targetm.gen_stack_protect_test (x, y, label); |
5012 | } |
5013 | |
5014 | if (seq) |
5015 | emit_insn (seq); |
5016 | else |
5017 | emit_cmp_and_jump_insns (x, y, EQ, NULL_RTX, ptr_mode, 1, label); |
5018 | |
5019 | /* The noreturn predictor has been moved to the tree level. The rtl-level |
5020 | predictors estimate this branch about 20%, which isn't enough to get |
5021 | things moved out of line. Since this is the only extant case of adding |
5022 | a noreturn function at the rtl level, it doesn't seem worth doing ought |
5023 | except adding the prediction by hand. */ |
5024 | rtx_insn *tmp = get_last_insn (); |
5025 | if (JUMP_P (tmp)) |
5026 | predict_insn_def (tmp, PRED_NORETURN, TAKEN); |
5027 | |
5028 | expand_call (targetm.stack_protect_fail (), NULL_RTX, /*ignore=*/true); |
5029 | free_temp_slots (); |
5030 | emit_label (label); |
5031 | } |
5032 | |
5033 | /* Start the RTL for a new function, and set variables used for |
5034 | emitting RTL. |
5035 | SUBR is the FUNCTION_DECL node. |
5036 | PARMS_HAVE_CLEANUPS is nonzero if there are cleanups associated with |
5037 | the function's parameters, which must be run at any return statement. */ |
5038 | |
5039 | bool currently_expanding_function_start; |
5040 | void |
5041 | expand_function_start (tree subr) |
5042 | { |
5043 | currently_expanding_function_start = true; |
5044 | |
5045 | /* Make sure volatile mem refs aren't considered |
5046 | valid operands of arithmetic insns. */ |
5047 | init_recog_no_volatile (); |
5048 | |
5049 | crtl->profile |
5050 | = (profile_flag |
5051 | && ! DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (subr)); |
5052 | |
5053 | crtl->limit_stack |
5054 | = (stack_limit_rtx != NULL_RTX && ! DECL_NO_LIMIT_STACK (subr)); |
5055 | |
5056 | /* Make the label for return statements to jump to. Do not special |
5057 | case machines with special return instructions -- they will be |
5058 | handled later during jump, ifcvt, or epilogue creation. */ |
5059 | return_label = gen_label_rtx (); |
5060 | |
5061 | /* Initialize rtx used to return the value. */ |
5062 | /* Do this before assign_parms so that we copy the struct value address |
5063 | before any library calls that assign parms might generate. */ |
5064 | |
5065 | /* Decide whether to return the value in memory or in a register. */ |
5066 | tree res = DECL_RESULT (subr); |
5067 | if (aggregate_value_p (exp: res, fntype: subr)) |
5068 | { |
5069 | /* Returning something that won't go in a register. */ |
5070 | rtx value_address = 0; |
5071 | |
5072 | #ifdef PCC_STATIC_STRUCT_RETURN |
5073 | if (cfun->returns_pcc_struct) |
5074 | { |
5075 | int size = int_size_in_bytes (TREE_TYPE (res)); |
5076 | value_address = assemble_static_space (size); |
5077 | } |
5078 | else |
5079 | #endif |
5080 | { |
5081 | rtx sv = targetm.calls.struct_value_rtx (TREE_TYPE (subr), 2); |
5082 | /* Expect to be passed the address of a place to store the value. |
5083 | If it is passed as an argument, assign_parms will take care of |
5084 | it. */ |
5085 | if (sv) |
5086 | { |
5087 | value_address = gen_reg_rtx (Pmode); |
5088 | emit_move_insn (value_address, sv); |
5089 | } |
5090 | } |
5091 | if (value_address) |
5092 | { |
5093 | rtx x = value_address; |
5094 | if (!DECL_BY_REFERENCE (res)) |
5095 | { |
5096 | x = gen_rtx_MEM (DECL_MODE (res), x); |
5097 | set_mem_attributes (x, res, 1); |
5098 | } |
5099 | set_parm_rtl (res, x); |
5100 | } |
5101 | } |
5102 | else if (DECL_MODE (res) == VOIDmode) |
5103 | /* If return mode is void, this decl rtl should not be used. */ |
5104 | set_parm_rtl (res, NULL_RTX); |
5105 | else |
5106 | { |
5107 | /* Compute the return values into a pseudo reg, which we will copy |
5108 | into the true return register after the cleanups are done. */ |
5109 | tree return_type = TREE_TYPE (res); |
5110 | |
5111 | /* If we may coalesce this result, make sure it has the expected mode |
5112 | in case it was promoted. But we need not bother about BLKmode. */ |
5113 | machine_mode promoted_mode |
5114 | = flag_tree_coalesce_vars && is_gimple_reg (res) |
5115 | ? promote_ssa_mode (ssa_default_def (cfun, res), NULL) |
5116 | : BLKmode; |
5117 | |
5118 | if (promoted_mode != BLKmode) |
5119 | set_parm_rtl (res, gen_reg_rtx (promoted_mode)); |
5120 | else if (TYPE_MODE (return_type) != BLKmode |
5121 | && targetm.calls.return_in_msb (return_type)) |
5122 | /* expand_function_end will insert the appropriate padding in |
5123 | this case. Use the return value's natural (unpadded) mode |
5124 | within the function proper. */ |
5125 | set_parm_rtl (res, gen_reg_rtx (TYPE_MODE (return_type))); |
5126 | else |
5127 | { |
5128 | /* In order to figure out what mode to use for the pseudo, we |
5129 | figure out what the mode of the eventual return register will |
5130 | actually be, and use that. */ |
5131 | rtx hard_reg = hard_function_value (return_type, subr, 0, 1); |
5132 | |
5133 | /* Structures that are returned in registers are not |
5134 | aggregate_value_p, so we may see a PARALLEL or a REG. */ |
5135 | if (REG_P (hard_reg)) |
5136 | set_parm_rtl (res, gen_reg_rtx (GET_MODE (hard_reg))); |
5137 | else |
5138 | { |
5139 | gcc_assert (GET_CODE (hard_reg) == PARALLEL); |
5140 | set_parm_rtl (res, gen_group_rtx (hard_reg)); |
5141 | } |
5142 | } |
5143 | |
5144 | /* Set DECL_REGISTER flag so that expand_function_end will copy the |
5145 | result to the real return register(s). */ |
5146 | DECL_REGISTER (res) = 1; |
5147 | } |
5148 | |
5149 | /* Initialize rtx for parameters and local variables. |
5150 | In some cases this requires emitting insns. */ |
5151 | assign_parms (fndecl: subr); |
5152 | |
5153 | /* If function gets a static chain arg, store it. */ |
5154 | if (cfun->static_chain_decl) |
5155 | { |
5156 | tree parm = cfun->static_chain_decl; |
5157 | rtx local, chain; |
5158 | rtx_insn *insn; |
5159 | int unsignedp; |
5160 | |
5161 | local = gen_reg_rtx (promote_decl_mode (parm, &unsignedp)); |
5162 | chain = targetm.calls.static_chain (current_function_decl, true); |
5163 | |
5164 | set_decl_incoming_rtl (parm, chain, false); |
5165 | set_parm_rtl (parm, local); |
5166 | mark_reg_pointer (local, TYPE_ALIGN (TREE_TYPE (TREE_TYPE (parm)))); |
5167 | |
5168 | if (GET_MODE (local) != GET_MODE (chain)) |
5169 | { |
5170 | convert_move (local, chain, unsignedp); |
5171 | insn = get_last_insn (); |
5172 | } |
5173 | else |
5174 | insn = emit_move_insn (local, chain); |
5175 | |
5176 | /* Mark the register as eliminable, similar to parameters. */ |
5177 | if (MEM_P (chain) |
5178 | && reg_mentioned_p (arg_pointer_rtx, XEXP (chain, 0))) |
5179 | set_dst_reg_note (insn, REG_EQUIV, chain, local); |
5180 | |
5181 | /* If we aren't optimizing, save the static chain onto the stack. */ |
5182 | if (!optimize) |
5183 | { |
5184 | tree saved_static_chain_decl |
5185 | = build_decl (DECL_SOURCE_LOCATION (parm), VAR_DECL, |
5186 | DECL_NAME (parm), TREE_TYPE (parm)); |
5187 | rtx saved_static_chain_rtx |
5188 | = assign_stack_local (Pmode, size: GET_MODE_SIZE (Pmode), align: 0); |
5189 | SET_DECL_RTL (saved_static_chain_decl, saved_static_chain_rtx); |
5190 | emit_move_insn (saved_static_chain_rtx, chain); |
5191 | SET_DECL_VALUE_EXPR (parm, saved_static_chain_decl); |
5192 | DECL_HAS_VALUE_EXPR_P (parm) = 1; |
5193 | } |
5194 | } |
5195 | |
5196 | /* The following was moved from init_function_start. |
5197 | The move was supposed to make sdb output more accurate. */ |
5198 | /* Indicate the beginning of the function body, |
5199 | as opposed to parm setup. */ |
5200 | emit_note (NOTE_INSN_FUNCTION_BEG); |
5201 | |
5202 | gcc_assert (NOTE_P (get_last_insn ())); |
5203 | |
5204 | parm_birth_insn = get_last_insn (); |
5205 | |
5206 | /* If the function receives a non-local goto, then store the |
5207 | bits we need to restore the frame pointer. */ |
5208 | if (cfun->nonlocal_goto_save_area) |
5209 | { |
5210 | tree t_save; |
5211 | rtx r_save; |
5212 | |
5213 | tree var = TREE_OPERAND (cfun->nonlocal_goto_save_area, 0); |
5214 | gcc_assert (DECL_RTL_SET_P (var)); |
5215 | |
5216 | t_save = build4 (ARRAY_REF, |
5217 | TREE_TYPE (TREE_TYPE (cfun->nonlocal_goto_save_area)), |
5218 | cfun->nonlocal_goto_save_area, |
5219 | integer_zero_node, NULL_TREE, NULL_TREE); |
5220 | r_save = expand_expr (exp: t_save, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
5221 | gcc_assert (GET_MODE (r_save) == Pmode); |
5222 | |
5223 | emit_move_insn (r_save, hard_frame_pointer_rtx); |
5224 | update_nonlocal_goto_save_area (); |
5225 | } |
5226 | |
5227 | if (crtl->profile) |
5228 | { |
5229 | #ifdef PROFILE_HOOK |
5230 | PROFILE_HOOK (current_function_funcdef_no); |
5231 | #endif |
5232 | } |
5233 | |
5234 | /* If we are doing generic stack checking, the probe should go here. */ |
5235 | if (flag_stack_check == GENERIC_STACK_CHECK) |
5236 | stack_check_probe_note = emit_note (NOTE_INSN_DELETED); |
5237 | |
5238 | currently_expanding_function_start = false; |
5239 | } |
5240 | |
5241 | void |
5242 | pop_dummy_function (void) |
5243 | { |
5244 | pop_cfun (); |
5245 | in_dummy_function = false; |
5246 | } |
5247 | |
5248 | /* Undo the effects of init_dummy_function_start. */ |
5249 | void |
5250 | expand_dummy_function_end (void) |
5251 | { |
5252 | gcc_assert (in_dummy_function); |
5253 | |
5254 | /* End any sequences that failed to be closed due to syntax errors. */ |
5255 | while (in_sequence_p ()) |
5256 | end_sequence (); |
5257 | |
5258 | /* Outside function body, can't compute type's actual size |
5259 | until next function's body starts. */ |
5260 | |
5261 | free_after_parsing (f: cfun); |
5262 | free_after_compilation (f: cfun); |
5263 | pop_dummy_function (); |
5264 | } |
5265 | |
5266 | /* Helper for diddle_return_value. */ |
5267 | |
5268 | void |
5269 | diddle_return_value_1 (void (*doit) (rtx, void *), void *arg, rtx outgoing) |
5270 | { |
5271 | if (! outgoing) |
5272 | return; |
5273 | |
5274 | if (REG_P (outgoing)) |
5275 | (*doit) (outgoing, arg); |
5276 | else if (GET_CODE (outgoing) == PARALLEL) |
5277 | { |
5278 | int i; |
5279 | |
5280 | for (i = 0; i < XVECLEN (outgoing, 0); i++) |
5281 | { |
5282 | rtx x = XEXP (XVECEXP (outgoing, 0, i), 0); |
5283 | |
5284 | if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER) |
5285 | (*doit) (x, arg); |
5286 | } |
5287 | } |
5288 | } |
5289 | |
5290 | /* Call DOIT for each hard register used as a return value from |
5291 | the current function. */ |
5292 | |
5293 | void |
5294 | diddle_return_value (void (*doit) (rtx, void *), void *arg) |
5295 | { |
5296 | diddle_return_value_1 (doit, arg, crtl->return_rtx); |
5297 | } |
5298 | |
5299 | static void |
5300 | do_clobber_return_reg (rtx reg, void *arg ATTRIBUTE_UNUSED) |
5301 | { |
5302 | emit_clobber (reg); |
5303 | } |
5304 | |
5305 | void |
5306 | clobber_return_register (void) |
5307 | { |
5308 | diddle_return_value (doit: do_clobber_return_reg, NULL); |
5309 | |
5310 | /* In case we do use pseudo to return value, clobber it too. */ |
5311 | if (DECL_RTL_SET_P (DECL_RESULT (current_function_decl))) |
5312 | { |
5313 | tree decl_result = DECL_RESULT (current_function_decl); |
5314 | rtx decl_rtl = DECL_RTL (decl_result); |
5315 | if (REG_P (decl_rtl) && REGNO (decl_rtl) >= FIRST_PSEUDO_REGISTER) |
5316 | { |
5317 | do_clobber_return_reg (reg: decl_rtl, NULL); |
5318 | } |
5319 | } |
5320 | } |
5321 | |
5322 | static void |
5323 | do_use_return_reg (rtx reg, void *arg ATTRIBUTE_UNUSED) |
5324 | { |
5325 | emit_use (reg); |
5326 | } |
5327 | |
5328 | static void |
5329 | use_return_register (void) |
5330 | { |
5331 | diddle_return_value (doit: do_use_return_reg, NULL); |
5332 | } |
5333 | |
5334 | /* Generate RTL for the end of the current function. */ |
5335 | |
5336 | void |
5337 | expand_function_end (void) |
5338 | { |
5339 | /* If arg_pointer_save_area was referenced only from a nested |
5340 | function, we will not have initialized it yet. Do that now. */ |
5341 | if (arg_pointer_save_area && ! crtl->arg_pointer_save_area_init) |
5342 | get_arg_pointer_save_area (); |
5343 | |
5344 | /* If we are doing generic stack checking and this function makes calls, |
5345 | do a stack probe at the start of the function to ensure we have enough |
5346 | space for another stack frame. */ |
5347 | if (flag_stack_check == GENERIC_STACK_CHECK) |
5348 | { |
5349 | rtx_insn *insn, *seq; |
5350 | |
5351 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
5352 | if (CALL_P (insn)) |
5353 | { |
5354 | rtx max_frame_size = GEN_INT (STACK_CHECK_MAX_FRAME_SIZE); |
5355 | start_sequence (); |
5356 | if (STACK_CHECK_MOVING_SP) |
5357 | anti_adjust_stack_and_probe (max_frame_size, true); |
5358 | else |
5359 | probe_stack_range (STACK_OLD_CHECK_PROTECT, max_frame_size); |
5360 | seq = get_insns (); |
5361 | end_sequence (); |
5362 | set_insn_locations (seq, prologue_location); |
5363 | emit_insn_before (seq, stack_check_probe_note); |
5364 | break; |
5365 | } |
5366 | } |
5367 | |
5368 | /* End any sequences that failed to be closed due to syntax errors. */ |
5369 | while (in_sequence_p ()) |
5370 | end_sequence (); |
5371 | |
5372 | clear_pending_stack_adjust (); |
5373 | do_pending_stack_adjust (); |
5374 | |
5375 | /* Output a linenumber for the end of the function. |
5376 | SDB depended on this. */ |
5377 | set_curr_insn_location (input_location); |
5378 | |
5379 | /* Before the return label (if any), clobber the return |
5380 | registers so that they are not propagated live to the rest of |
5381 | the function. This can only happen with functions that drop |
5382 | through; if there had been a return statement, there would |
5383 | have either been a return rtx, or a jump to the return label. |
5384 | |
5385 | We delay actual code generation after the current_function_value_rtx |
5386 | is computed. */ |
5387 | rtx_insn *clobber_after = get_last_insn (); |
5388 | |
5389 | /* Output the label for the actual return from the function. */ |
5390 | emit_label (return_label); |
5391 | |
5392 | if (targetm_common.except_unwind_info (&global_options) == UI_SJLJ) |
5393 | { |
5394 | /* Let except.cc know where it should emit the call to unregister |
5395 | the function context for sjlj exceptions. */ |
5396 | if (flag_exceptions) |
5397 | sjlj_emit_function_exit_after (get_last_insn ()); |
5398 | } |
5399 | |
5400 | /* If this is an implementation of throw, do what's necessary to |
5401 | communicate between __builtin_eh_return and the epilogue. */ |
5402 | expand_eh_return (); |
5403 | |
5404 | /* If stack protection is enabled for this function, check the guard. */ |
5405 | if (crtl->stack_protect_guard |
5406 | && targetm.stack_protect_runtime_enabled_p () |
5407 | && naked_return_label == NULL_RTX) |
5408 | stack_protect_epilogue (); |
5409 | |
5410 | /* If scalar return value was computed in a pseudo-reg, or was a named |
5411 | return value that got dumped to the stack, copy that to the hard |
5412 | return register. */ |
5413 | if (DECL_RTL_SET_P (DECL_RESULT (current_function_decl))) |
5414 | { |
5415 | tree decl_result = DECL_RESULT (current_function_decl); |
5416 | rtx decl_rtl = DECL_RTL (decl_result); |
5417 | |
5418 | if ((REG_P (decl_rtl) |
5419 | ? REGNO (decl_rtl) >= FIRST_PSEUDO_REGISTER |
5420 | : DECL_REGISTER (decl_result)) |
5421 | /* Unless the psABI says not to. */ |
5422 | && !TYPE_EMPTY_P (TREE_TYPE (decl_result))) |
5423 | { |
5424 | rtx real_decl_rtl = crtl->return_rtx; |
5425 | complex_mode cmode; |
5426 | |
5427 | /* This should be set in assign_parms. */ |
5428 | gcc_assert (REG_FUNCTION_VALUE_P (real_decl_rtl)); |
5429 | |
5430 | /* If this is a BLKmode structure being returned in registers, |
5431 | then use the mode computed in expand_return. Note that if |
5432 | decl_rtl is memory, then its mode may have been changed, |
5433 | but that crtl->return_rtx has not. */ |
5434 | if (GET_MODE (real_decl_rtl) == BLKmode) |
5435 | PUT_MODE (x: real_decl_rtl, GET_MODE (decl_rtl)); |
5436 | |
5437 | /* If a non-BLKmode return value should be padded at the least |
5438 | significant end of the register, shift it left by the appropriate |
5439 | amount. BLKmode results are handled using the group load/store |
5440 | machinery. */ |
5441 | if (TYPE_MODE (TREE_TYPE (decl_result)) != BLKmode |
5442 | && REG_P (real_decl_rtl) |
5443 | && targetm.calls.return_in_msb (TREE_TYPE (decl_result))) |
5444 | { |
5445 | emit_move_insn (gen_rtx_REG (GET_MODE (decl_rtl), |
5446 | REGNO (real_decl_rtl)), |
5447 | decl_rtl); |
5448 | shift_return_value (GET_MODE (decl_rtl), true, real_decl_rtl); |
5449 | } |
5450 | else if (GET_CODE (real_decl_rtl) == PARALLEL) |
5451 | { |
5452 | /* If expand_function_start has created a PARALLEL for decl_rtl, |
5453 | move the result to the real return registers. Otherwise, do |
5454 | a group load from decl_rtl for a named return. */ |
5455 | if (GET_CODE (decl_rtl) == PARALLEL) |
5456 | emit_group_move (real_decl_rtl, decl_rtl); |
5457 | else |
5458 | emit_group_load (real_decl_rtl, decl_rtl, |
5459 | TREE_TYPE (decl_result), |
5460 | int_size_in_bytes (TREE_TYPE (decl_result))); |
5461 | } |
5462 | /* In the case of complex integer modes smaller than a word, we'll |
5463 | need to generate some non-trivial bitfield insertions. Do that |
5464 | on a pseudo and not the hard register. */ |
5465 | else if (GET_CODE (decl_rtl) == CONCAT |
5466 | && is_complex_int_mode (GET_MODE (decl_rtl), cmode: &cmode) |
5467 | && GET_MODE_BITSIZE (mode: cmode) <= BITS_PER_WORD) |
5468 | { |
5469 | int old_generating_concat_p; |
5470 | rtx tmp; |
5471 | |
5472 | old_generating_concat_p = generating_concat_p; |
5473 | generating_concat_p = 0; |
5474 | tmp = gen_reg_rtx (GET_MODE (decl_rtl)); |
5475 | generating_concat_p = old_generating_concat_p; |
5476 | |
5477 | emit_move_insn (tmp, decl_rtl); |
5478 | emit_move_insn (real_decl_rtl, tmp); |
5479 | } |
5480 | /* If a named return value dumped decl_return to memory, then |
5481 | we may need to re-do the PROMOTE_MODE signed/unsigned |
5482 | extension. */ |
5483 | else if (GET_MODE (real_decl_rtl) != GET_MODE (decl_rtl)) |
5484 | { |
5485 | int unsignedp = TYPE_UNSIGNED (TREE_TYPE (decl_result)); |
5486 | promote_function_mode (TREE_TYPE (decl_result), |
5487 | GET_MODE (decl_rtl), &unsignedp, |
5488 | TREE_TYPE (current_function_decl), 1); |
5489 | |
5490 | convert_move (real_decl_rtl, decl_rtl, unsignedp); |
5491 | } |
5492 | else |
5493 | emit_move_insn (real_decl_rtl, decl_rtl); |
5494 | } |
5495 | } |
5496 | |
5497 | /* If returning a structure, arrange to return the address of the value |
5498 | in a place where debuggers expect to find it. |
5499 | |
5500 | If returning a structure PCC style, |
5501 | the caller also depends on this value. |
5502 | And cfun->returns_pcc_struct is not necessarily set. */ |
5503 | if ((cfun->returns_struct || cfun->returns_pcc_struct) |
5504 | && !targetm.calls.omit_struct_return_reg) |
5505 | { |
5506 | rtx value_address = DECL_RTL (DECL_RESULT (current_function_decl)); |
5507 | tree type = TREE_TYPE (DECL_RESULT (current_function_decl)); |
5508 | rtx outgoing; |
5509 | |
5510 | if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) |
5511 | type = TREE_TYPE (type); |
5512 | else |
5513 | value_address = XEXP (value_address, 0); |
5514 | |
5515 | outgoing = targetm.calls.function_value (build_pointer_type (type), |
5516 | current_function_decl, true); |
5517 | |
5518 | /* Mark this as a function return value so integrate will delete the |
5519 | assignment and USE below when inlining this function. */ |
5520 | REG_FUNCTION_VALUE_P (outgoing) = 1; |
5521 | |
5522 | /* The address may be ptr_mode and OUTGOING may be Pmode. */ |
5523 | scalar_int_mode mode = as_a <scalar_int_mode> (GET_MODE (outgoing)); |
5524 | value_address = convert_memory_address (mode, value_address); |
5525 | |
5526 | emit_move_insn (outgoing, value_address); |
5527 | |
5528 | /* Show return register used to hold result (in this case the address |
5529 | of the result. */ |
5530 | crtl->return_rtx = outgoing; |
5531 | } |
5532 | |
5533 | /* Emit the actual code to clobber return register. Don't emit |
5534 | it if clobber_after is a barrier, then the previous basic block |
5535 | certainly doesn't fall thru into the exit block. */ |
5536 | if (!BARRIER_P (clobber_after)) |
5537 | { |
5538 | start_sequence (); |
5539 | clobber_return_register (); |
5540 | rtx_insn *seq = get_insns (); |
5541 | end_sequence (); |
5542 | |
5543 | emit_insn_after (seq, clobber_after); |
5544 | } |
5545 | |
5546 | /* Output the label for the naked return from the function. */ |
5547 | if (naked_return_label) |
5548 | emit_label (naked_return_label); |
5549 | |
5550 | /* @@@ This is a kludge. We want to ensure that instructions that |
5551 | may trap are not moved into the epilogue by scheduling, because |
5552 | we don't always emit unwind information for the epilogue. */ |
5553 | if (cfun->can_throw_non_call_exceptions |
5554 | && targetm_common.except_unwind_info (&global_options) != UI_SJLJ) |
5555 | emit_insn (gen_blockage ()); |
5556 | |
5557 | /* If stack protection is enabled for this function, check the guard. */ |
5558 | if (crtl->stack_protect_guard |
5559 | && targetm.stack_protect_runtime_enabled_p () |
5560 | && naked_return_label) |
5561 | stack_protect_epilogue (); |
5562 | |
5563 | /* If we had calls to alloca, and this machine needs |
5564 | an accurate stack pointer to exit the function, |
5565 | insert some code to save and restore the stack pointer. */ |
5566 | if (! EXIT_IGNORE_STACK |
5567 | && cfun->calls_alloca) |
5568 | { |
5569 | rtx tem = 0; |
5570 | |
5571 | start_sequence (); |
5572 | emit_stack_save (SAVE_FUNCTION, &tem); |
5573 | rtx_insn *seq = get_insns (); |
5574 | end_sequence (); |
5575 | emit_insn_before (seq, parm_birth_insn); |
5576 | |
5577 | emit_stack_restore (SAVE_FUNCTION, tem); |
5578 | } |
5579 | |
5580 | /* ??? This should no longer be necessary since stupid is no longer with |
5581 | us, but there are some parts of the compiler (eg reload_combine, and |
5582 | sh mach_dep_reorg) that still try and compute their own lifetime info |
5583 | instead of using the general framework. */ |
5584 | use_return_register (); |
5585 | } |
5586 | |
5587 | rtx |
5588 | get_arg_pointer_save_area (void) |
5589 | { |
5590 | rtx ret = arg_pointer_save_area; |
5591 | |
5592 | if (! ret) |
5593 | { |
5594 | ret = assign_stack_local (Pmode, size: GET_MODE_SIZE (Pmode), align: 0); |
5595 | arg_pointer_save_area = ret; |
5596 | } |
5597 | |
5598 | if (! crtl->arg_pointer_save_area_init) |
5599 | { |
5600 | /* Save the arg pointer at the beginning of the function. The |
5601 | generated stack slot may not be a valid memory address, so we |
5602 | have to check it and fix it if necessary. */ |
5603 | start_sequence (); |
5604 | emit_move_insn (validize_mem (copy_rtx (ret)), |
5605 | crtl->args.internal_arg_pointer); |
5606 | rtx_insn *seq = get_insns (); |
5607 | end_sequence (); |
5608 | |
5609 | push_topmost_sequence (); |
5610 | emit_insn_after (seq, entry_of_function ()); |
5611 | pop_topmost_sequence (); |
5612 | |
5613 | crtl->arg_pointer_save_area_init = true; |
5614 | } |
5615 | |
5616 | return ret; |
5617 | } |
5618 | |
5619 | |
5620 | /* If debugging dumps are requested, dump information about how the |
5621 | target handled -fstack-check=clash for the prologue. |
5622 | |
5623 | PROBES describes what if any probes were emitted. |
5624 | |
5625 | RESIDUALS indicates if the prologue had any residual allocation |
5626 | (i.e. total allocation was not a multiple of PROBE_INTERVAL). */ |
5627 | |
5628 | void |
5629 | dump_stack_clash_frame_info (enum stack_clash_probes probes, bool residuals) |
5630 | { |
5631 | if (!dump_file) |
5632 | return; |
5633 | |
5634 | switch (probes) |
5635 | { |
5636 | case NO_PROBE_NO_FRAME: |
5637 | fprintf (stream: dump_file, |
5638 | format: "Stack clash no probe no stack adjustment in prologue.\n" ); |
5639 | break; |
5640 | case NO_PROBE_SMALL_FRAME: |
5641 | fprintf (stream: dump_file, |
5642 | format: "Stack clash no probe small stack adjustment in prologue.\n" ); |
5643 | break; |
5644 | case PROBE_INLINE: |
5645 | fprintf (stream: dump_file, format: "Stack clash inline probes in prologue.\n" ); |
5646 | break; |
5647 | case PROBE_LOOP: |
5648 | fprintf (stream: dump_file, format: "Stack clash probe loop in prologue.\n" ); |
5649 | break; |
5650 | } |
5651 | |
5652 | if (residuals) |
5653 | fprintf (stream: dump_file, format: "Stack clash residual allocation in prologue.\n" ); |
5654 | else |
5655 | fprintf (stream: dump_file, format: "Stack clash no residual allocation in prologue.\n" ); |
5656 | |
5657 | if (frame_pointer_needed) |
5658 | fprintf (stream: dump_file, format: "Stack clash frame pointer needed.\n" ); |
5659 | else |
5660 | fprintf (stream: dump_file, format: "Stack clash no frame pointer needed.\n" ); |
5661 | |
5662 | if (TREE_THIS_VOLATILE (cfun->decl)) |
5663 | fprintf (stream: dump_file, |
5664 | format: "Stack clash noreturn prologue, assuming no implicit" |
5665 | " probes in caller.\n" ); |
5666 | else |
5667 | fprintf (stream: dump_file, |
5668 | format: "Stack clash not noreturn prologue.\n" ); |
5669 | } |
5670 | |
5671 | /* Add a list of INSNS to the hash HASHP, possibly allocating HASHP |
5672 | for the first time. */ |
5673 | |
5674 | static void |
5675 | record_insns (rtx_insn *insns, rtx end, hash_table<insn_cache_hasher> **hashp) |
5676 | { |
5677 | rtx_insn *tmp; |
5678 | hash_table<insn_cache_hasher> *hash = *hashp; |
5679 | |
5680 | if (hash == NULL) |
5681 | *hashp = hash = hash_table<insn_cache_hasher>::create_ggc (n: 17); |
5682 | |
5683 | for (tmp = insns; tmp != end; tmp = NEXT_INSN (insn: tmp)) |
5684 | { |
5685 | rtx *slot = hash->find_slot (value: tmp, insert: INSERT); |
5686 | gcc_assert (*slot == NULL); |
5687 | *slot = tmp; |
5688 | } |
5689 | } |
5690 | |
5691 | /* INSN has been duplicated or replaced by as COPY, perhaps by duplicating a |
5692 | basic block, splitting or peepholes. If INSN is a prologue or epilogue |
5693 | insn, then record COPY as well. */ |
5694 | |
5695 | void |
5696 | maybe_copy_prologue_epilogue_insn (rtx insn, rtx copy) |
5697 | { |
5698 | hash_table<insn_cache_hasher> *hash; |
5699 | rtx *slot; |
5700 | |
5701 | hash = epilogue_insn_hash; |
5702 | if (!hash || !hash->find (value: insn)) |
5703 | { |
5704 | hash = prologue_insn_hash; |
5705 | if (!hash || !hash->find (value: insn)) |
5706 | return; |
5707 | } |
5708 | |
5709 | slot = hash->find_slot (value: copy, insert: INSERT); |
5710 | gcc_assert (*slot == NULL); |
5711 | *slot = copy; |
5712 | } |
5713 | |
5714 | /* Determine if any INSNs in HASH are, or are part of, INSN. Because |
5715 | we can be running after reorg, SEQUENCE rtl is possible. */ |
5716 | |
5717 | static bool |
5718 | contains (const rtx_insn *insn, hash_table<insn_cache_hasher> *hash) |
5719 | { |
5720 | if (hash == NULL) |
5721 | return false; |
5722 | |
5723 | if (NONJUMP_INSN_P (insn) && GET_CODE (PATTERN (insn)) == SEQUENCE) |
5724 | { |
5725 | rtx_sequence *seq = as_a <rtx_sequence *> (p: PATTERN (insn)); |
5726 | int i; |
5727 | for (i = seq->len () - 1; i >= 0; i--) |
5728 | if (hash->find (value: seq->element (index: i))) |
5729 | return true; |
5730 | return false; |
5731 | } |
5732 | |
5733 | return hash->find (value: const_cast<rtx_insn *> (insn)) != NULL; |
5734 | } |
5735 | |
5736 | bool |
5737 | prologue_contains (const rtx_insn *insn) |
5738 | { |
5739 | return contains (insn, hash: prologue_insn_hash); |
5740 | } |
5741 | |
5742 | bool |
5743 | epilogue_contains (const rtx_insn *insn) |
5744 | { |
5745 | return contains (insn, hash: epilogue_insn_hash); |
5746 | } |
5747 | |
5748 | bool |
5749 | prologue_epilogue_contains (const rtx_insn *insn) |
5750 | { |
5751 | if (contains (insn, hash: prologue_insn_hash)) |
5752 | return true; |
5753 | if (contains (insn, hash: epilogue_insn_hash)) |
5754 | return true; |
5755 | return false; |
5756 | } |
5757 | |
5758 | void |
5759 | record_prologue_seq (rtx_insn *seq) |
5760 | { |
5761 | record_insns (insns: seq, NULL, hashp: &prologue_insn_hash); |
5762 | } |
5763 | |
5764 | void |
5765 | record_epilogue_seq (rtx_insn *seq) |
5766 | { |
5767 | record_insns (insns: seq, NULL, hashp: &epilogue_insn_hash); |
5768 | } |
5769 | |
5770 | /* Set JUMP_LABEL for a return insn. */ |
5771 | |
5772 | void |
5773 | set_return_jump_label (rtx_insn *returnjump) |
5774 | { |
5775 | rtx pat = PATTERN (insn: returnjump); |
5776 | if (GET_CODE (pat) == PARALLEL) |
5777 | pat = XVECEXP (pat, 0, 0); |
5778 | if (ANY_RETURN_P (pat)) |
5779 | JUMP_LABEL (returnjump) = pat; |
5780 | else |
5781 | JUMP_LABEL (returnjump) = ret_rtx; |
5782 | } |
5783 | |
5784 | /* Return a sequence to be used as the split prologue for the current |
5785 | function, or NULL. */ |
5786 | |
5787 | static rtx_insn * |
5788 | make_split_prologue_seq (void) |
5789 | { |
5790 | if (!flag_split_stack |
5791 | || lookup_attribute (attr_name: "no_split_stack" , DECL_ATTRIBUTES (cfun->decl))) |
5792 | return NULL; |
5793 | |
5794 | start_sequence (); |
5795 | emit_insn (targetm.gen_split_stack_prologue ()); |
5796 | rtx_insn *seq = get_insns (); |
5797 | end_sequence (); |
5798 | |
5799 | record_insns (insns: seq, NULL, hashp: &prologue_insn_hash); |
5800 | set_insn_locations (seq, prologue_location); |
5801 | |
5802 | return seq; |
5803 | } |
5804 | |
5805 | /* Return a sequence to be used as the prologue for the current function, |
5806 | or NULL. */ |
5807 | |
5808 | static rtx_insn * |
5809 | make_prologue_seq (void) |
5810 | { |
5811 | if (!targetm.have_prologue ()) |
5812 | return NULL; |
5813 | |
5814 | start_sequence (); |
5815 | rtx_insn *seq = targetm.gen_prologue (); |
5816 | emit_insn (seq); |
5817 | |
5818 | /* Insert an explicit USE for the frame pointer |
5819 | if the profiling is on and the frame pointer is required. */ |
5820 | if (crtl->profile && frame_pointer_needed) |
5821 | emit_use (hard_frame_pointer_rtx); |
5822 | |
5823 | /* Retain a map of the prologue insns. */ |
5824 | record_insns (insns: seq, NULL, hashp: &prologue_insn_hash); |
5825 | emit_note (NOTE_INSN_PROLOGUE_END); |
5826 | |
5827 | /* Ensure that instructions are not moved into the prologue when |
5828 | profiling is on. The call to the profiling routine can be |
5829 | emitted within the live range of a call-clobbered register. */ |
5830 | if (!targetm.profile_before_prologue () && crtl->profile) |
5831 | emit_insn (gen_blockage ()); |
5832 | |
5833 | seq = get_insns (); |
5834 | end_sequence (); |
5835 | set_insn_locations (seq, prologue_location); |
5836 | |
5837 | return seq; |
5838 | } |
5839 | |
5840 | /* Emit a sequence of insns to zero the call-used registers before RET |
5841 | according to ZERO_REGS_TYPE. */ |
5842 | |
5843 | static void |
5844 | gen_call_used_regs_seq (rtx_insn *ret, unsigned int zero_regs_type) |
5845 | { |
5846 | bool only_gpr = true; |
5847 | bool only_used = true; |
5848 | bool only_arg = true; |
5849 | |
5850 | /* No need to zero call-used-regs in main (). */ |
5851 | if (MAIN_NAME_P (DECL_NAME (current_function_decl))) |
5852 | return; |
5853 | |
5854 | /* No need to zero call-used-regs if __builtin_eh_return is called |
5855 | since it isn't a normal function return. */ |
5856 | if (crtl->calls_eh_return) |
5857 | return; |
5858 | |
5859 | /* If only_gpr is true, only zero call-used registers that are |
5860 | general-purpose registers; if only_used is true, only zero |
5861 | call-used registers that are used in the current function; |
5862 | if only_arg is true, only zero call-used registers that pass |
5863 | parameters defined by the flatform's calling conversion. */ |
5864 | |
5865 | using namespace zero_regs_flags; |
5866 | |
5867 | only_gpr = zero_regs_type & ONLY_GPR; |
5868 | only_used = zero_regs_type & ONLY_USED; |
5869 | only_arg = zero_regs_type & ONLY_ARG; |
5870 | |
5871 | if ((zero_regs_type & LEAFY_MODE) && leaf_function_p ()) |
5872 | only_used = true; |
5873 | |
5874 | /* For each of the hard registers, we should zero it if: |
5875 | 1. it is a call-used register; |
5876 | and 2. it is not a fixed register; |
5877 | and 3. it is not live at the return of the routine; |
5878 | and 4. it is general registor if only_gpr is true; |
5879 | and 5. it is used in the routine if only_used is true; |
5880 | and 6. it is a register that passes parameter if only_arg is true. */ |
5881 | |
5882 | /* First, prepare the data flow information. */ |
5883 | basic_block bb = BLOCK_FOR_INSN (insn: ret); |
5884 | auto_bitmap live_out; |
5885 | bitmap_copy (live_out, df_get_live_out (bb)); |
5886 | df_simulate_initialize_backwards (bb, live_out); |
5887 | df_simulate_one_insn_backwards (bb, ret, live_out); |
5888 | |
5889 | HARD_REG_SET selected_hardregs; |
5890 | HARD_REG_SET all_call_used_regs; |
5891 | CLEAR_HARD_REG_SET (set&: selected_hardregs); |
5892 | CLEAR_HARD_REG_SET (set&: all_call_used_regs); |
5893 | for (unsigned int regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) |
5894 | { |
5895 | if (!crtl->abi->clobbers_full_reg_p (regno)) |
5896 | continue; |
5897 | if (fixed_regs[regno]) |
5898 | continue; |
5899 | if (REGNO_REG_SET_P (live_out, regno)) |
5900 | continue; |
5901 | #ifdef LEAF_REG_REMAP |
5902 | if (crtl->uses_only_leaf_regs && LEAF_REG_REMAP (regno) < 0) |
5903 | continue; |
5904 | #endif |
5905 | /* This is a call used register that is dead at return. */ |
5906 | SET_HARD_REG_BIT (set&: all_call_used_regs, bit: regno); |
5907 | |
5908 | if (only_gpr |
5909 | && !TEST_HARD_REG_BIT (reg_class_contents[GENERAL_REGS], bit: regno)) |
5910 | continue; |
5911 | if (only_used && !df_regs_ever_live_p (regno)) |
5912 | continue; |
5913 | if (only_arg && !FUNCTION_ARG_REGNO_P (regno)) |
5914 | continue; |
5915 | |
5916 | /* Now this is a register that we might want to zero. */ |
5917 | SET_HARD_REG_BIT (set&: selected_hardregs, bit: regno); |
5918 | } |
5919 | |
5920 | if (hard_reg_set_empty_p (x: selected_hardregs)) |
5921 | return; |
5922 | |
5923 | /* Now that we have a hard register set that needs to be zeroed, pass it to |
5924 | target to generate zeroing sequence. */ |
5925 | HARD_REG_SET zeroed_hardregs; |
5926 | start_sequence (); |
5927 | zeroed_hardregs = targetm.calls.zero_call_used_regs (selected_hardregs); |
5928 | |
5929 | /* For most targets, the returned set of registers is a subset of |
5930 | selected_hardregs, however, for some of the targets (for example MIPS), |
5931 | clearing some registers that are in selected_hardregs requires clearing |
5932 | other call used registers that are not in the selected_hardregs, under |
5933 | such situation, the returned set of registers must be a subset of |
5934 | all call used registers. */ |
5935 | gcc_assert (hard_reg_set_subset_p (zeroed_hardregs, all_call_used_regs)); |
5936 | |
5937 | rtx_insn *seq = get_insns (); |
5938 | end_sequence (); |
5939 | if (seq) |
5940 | { |
5941 | /* Emit the memory blockage and register clobber asm volatile before |
5942 | the whole sequence. */ |
5943 | start_sequence (); |
5944 | expand_asm_reg_clobber_mem_blockage (zeroed_hardregs); |
5945 | rtx_insn *seq_barrier = get_insns (); |
5946 | end_sequence (); |
5947 | |
5948 | emit_insn_before (seq_barrier, ret); |
5949 | emit_insn_before (seq, ret); |
5950 | |
5951 | /* Update the data flow information. */ |
5952 | crtl->must_be_zero_on_return |= zeroed_hardregs; |
5953 | df_update_exit_block_uses (); |
5954 | } |
5955 | } |
5956 | |
5957 | |
5958 | /* Return a sequence to be used as the epilogue for the current function, |
5959 | or NULL. */ |
5960 | |
5961 | static rtx_insn * |
5962 | make_epilogue_seq (void) |
5963 | { |
5964 | if (!targetm.have_epilogue ()) |
5965 | return NULL; |
5966 | |
5967 | start_sequence (); |
5968 | emit_note (NOTE_INSN_EPILOGUE_BEG); |
5969 | rtx_insn *seq = targetm.gen_epilogue (); |
5970 | if (seq) |
5971 | emit_jump_insn (seq); |
5972 | |
5973 | /* Retain a map of the epilogue insns. */ |
5974 | record_insns (insns: seq, NULL, hashp: &epilogue_insn_hash); |
5975 | set_insn_locations (seq, epilogue_location); |
5976 | |
5977 | seq = get_insns (); |
5978 | rtx_insn *returnjump = get_last_insn (); |
5979 | end_sequence (); |
5980 | |
5981 | if (JUMP_P (returnjump)) |
5982 | set_return_jump_label (returnjump); |
5983 | |
5984 | return seq; |
5985 | } |
5986 | |
5987 | |
5988 | /* Generate the prologue and epilogue RTL if the machine supports it. Thread |
5989 | this into place with notes indicating where the prologue ends and where |
5990 | the epilogue begins. Update the basic block information when possible. |
5991 | |
5992 | Notes on epilogue placement: |
5993 | There are several kinds of edges to the exit block: |
5994 | * a single fallthru edge from LAST_BB |
5995 | * possibly, edges from blocks containing sibcalls |
5996 | * possibly, fake edges from infinite loops |
5997 | |
5998 | The epilogue is always emitted on the fallthru edge from the last basic |
5999 | block in the function, LAST_BB, into the exit block. |
6000 | |
6001 | If LAST_BB is empty except for a label, it is the target of every |
6002 | other basic block in the function that ends in a return. If a |
6003 | target has a return or simple_return pattern (possibly with |
6004 | conditional variants), these basic blocks can be changed so that a |
6005 | return insn is emitted into them, and their target is adjusted to |
6006 | the real exit block. |
6007 | |
6008 | Notes on shrink wrapping: We implement a fairly conservative |
6009 | version of shrink-wrapping rather than the textbook one. We only |
6010 | generate a single prologue and a single epilogue. This is |
6011 | sufficient to catch a number of interesting cases involving early |
6012 | exits. |
6013 | |
6014 | First, we identify the blocks that require the prologue to occur before |
6015 | them. These are the ones that modify a call-saved register, or reference |
6016 | any of the stack or frame pointer registers. To simplify things, we then |
6017 | mark everything reachable from these blocks as also requiring a prologue. |
6018 | This takes care of loops automatically, and avoids the need to examine |
6019 | whether MEMs reference the frame, since it is sufficient to check for |
6020 | occurrences of the stack or frame pointer. |
6021 | |
6022 | We then compute the set of blocks for which the need for a prologue |
6023 | is anticipatable (borrowing terminology from the shrink-wrapping |
6024 | description in Muchnick's book). These are the blocks which either |
6025 | require a prologue themselves, or those that have only successors |
6026 | where the prologue is anticipatable. The prologue needs to be |
6027 | inserted on all edges from BB1->BB2 where BB2 is in ANTIC and BB1 |
6028 | is not. For the moment, we ensure that only one such edge exists. |
6029 | |
6030 | The epilogue is placed as described above, but we make a |
6031 | distinction between inserting return and simple_return patterns |
6032 | when modifying other blocks that end in a return. Blocks that end |
6033 | in a sibcall omit the sibcall_epilogue if the block is not in |
6034 | ANTIC. */ |
6035 | |
6036 | void |
6037 | thread_prologue_and_epilogue_insns (void) |
6038 | { |
6039 | df_analyze (); |
6040 | |
6041 | /* Can't deal with multiple successors of the entry block at the |
6042 | moment. Function should always have at least one entry |
6043 | point. */ |
6044 | gcc_assert (single_succ_p (ENTRY_BLOCK_PTR_FOR_FN (cfun))); |
6045 | |
6046 | edge entry_edge = single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
6047 | edge orig_entry_edge = entry_edge; |
6048 | |
6049 | rtx_insn *split_prologue_seq = make_split_prologue_seq (); |
6050 | rtx_insn *prologue_seq = make_prologue_seq (); |
6051 | rtx_insn *epilogue_seq = make_epilogue_seq (); |
6052 | |
6053 | /* Try to perform a kind of shrink-wrapping, making sure the |
6054 | prologue/epilogue is emitted only around those parts of the |
6055 | function that require it. */ |
6056 | try_shrink_wrapping (entry_edge: &entry_edge, prologue_seq); |
6057 | |
6058 | /* If the target can handle splitting the prologue/epilogue into separate |
6059 | components, try to shrink-wrap these components separately. */ |
6060 | try_shrink_wrapping_separate (first_bb: entry_edge->dest); |
6061 | |
6062 | /* If that did anything for any component we now need the generate the |
6063 | "main" prologue again. Because some targets require some of these |
6064 | to be called in a specific order (i386 requires the split prologue |
6065 | to be first, for example), we create all three sequences again here. |
6066 | If this does not work for some target, that target should not enable |
6067 | separate shrink-wrapping. */ |
6068 | if (crtl->shrink_wrapped_separate) |
6069 | { |
6070 | split_prologue_seq = make_split_prologue_seq (); |
6071 | prologue_seq = make_prologue_seq (); |
6072 | epilogue_seq = make_epilogue_seq (); |
6073 | } |
6074 | |
6075 | rtl_profile_for_bb (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
6076 | |
6077 | /* A small fib -- epilogue is not yet completed, but we wish to re-use |
6078 | this marker for the splits of EH_RETURN patterns, and nothing else |
6079 | uses the flag in the meantime. */ |
6080 | epilogue_completed = 1; |
6081 | |
6082 | /* Find non-fallthru edges that end with EH_RETURN instructions. On |
6083 | some targets, these get split to a special version of the epilogue |
6084 | code. In order to be able to properly annotate these with unwind |
6085 | info, try to split them now. If we get a valid split, drop an |
6086 | EPILOGUE_BEG note and mark the insns as epilogue insns. */ |
6087 | edge e; |
6088 | edge_iterator ei; |
6089 | FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) |
6090 | { |
6091 | rtx_insn *prev, *last, *trial; |
6092 | |
6093 | if (e->flags & EDGE_FALLTHRU) |
6094 | continue; |
6095 | last = BB_END (e->src); |
6096 | if (!eh_returnjump_p (last)) |
6097 | continue; |
6098 | |
6099 | prev = PREV_INSN (insn: last); |
6100 | trial = try_split (PATTERN (insn: last), last, 1); |
6101 | if (trial == last) |
6102 | continue; |
6103 | |
6104 | record_insns (insns: NEXT_INSN (insn: prev), end: NEXT_INSN (insn: trial), hashp: &epilogue_insn_hash); |
6105 | emit_note_after (NOTE_INSN_EPILOGUE_BEG, prev); |
6106 | } |
6107 | |
6108 | edge exit_fallthru_edge = find_fallthru_edge (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds); |
6109 | |
6110 | if (exit_fallthru_edge) |
6111 | { |
6112 | if (epilogue_seq) |
6113 | { |
6114 | insert_insn_on_edge (epilogue_seq, exit_fallthru_edge); |
6115 | commit_edge_insertions (); |
6116 | |
6117 | /* The epilogue insns we inserted may cause the exit edge to no longer |
6118 | be fallthru. */ |
6119 | FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) |
6120 | { |
6121 | if (((e->flags & EDGE_FALLTHRU) != 0) |
6122 | && returnjump_p (BB_END (e->src))) |
6123 | e->flags &= ~EDGE_FALLTHRU; |
6124 | } |
6125 | |
6126 | find_sub_basic_blocks (BLOCK_FOR_INSN (insn: epilogue_seq)); |
6127 | } |
6128 | else if (next_active_insn (BB_END (exit_fallthru_edge->src))) |
6129 | { |
6130 | /* We have a fall-through edge to the exit block, the source is not |
6131 | at the end of the function, and there will be an assembler epilogue |
6132 | at the end of the function. |
6133 | We can't use force_nonfallthru here, because that would try to |
6134 | use return. Inserting a jump 'by hand' is extremely messy, so |
6135 | we take advantage of cfg_layout_finalize using |
6136 | fixup_fallthru_exit_predecessor. */ |
6137 | cfg_layout_initialize (0); |
6138 | basic_block cur_bb; |
6139 | FOR_EACH_BB_FN (cur_bb, cfun) |
6140 | if (cur_bb->index >= NUM_FIXED_BLOCKS |
6141 | && cur_bb->next_bb->index >= NUM_FIXED_BLOCKS) |
6142 | cur_bb->aux = cur_bb->next_bb; |
6143 | cfg_layout_finalize (); |
6144 | } |
6145 | } |
6146 | |
6147 | /* Insert the prologue. */ |
6148 | |
6149 | rtl_profile_for_bb (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
6150 | |
6151 | if (split_prologue_seq || prologue_seq) |
6152 | { |
6153 | rtx_insn *split_prologue_insn = split_prologue_seq; |
6154 | if (split_prologue_seq) |
6155 | { |
6156 | while (split_prologue_insn && !NONDEBUG_INSN_P (split_prologue_insn)) |
6157 | split_prologue_insn = NEXT_INSN (insn: split_prologue_insn); |
6158 | insert_insn_on_edge (split_prologue_seq, orig_entry_edge); |
6159 | } |
6160 | |
6161 | rtx_insn *prologue_insn = prologue_seq; |
6162 | if (prologue_seq) |
6163 | { |
6164 | while (prologue_insn && !NONDEBUG_INSN_P (prologue_insn)) |
6165 | prologue_insn = NEXT_INSN (insn: prologue_insn); |
6166 | insert_insn_on_edge (prologue_seq, entry_edge); |
6167 | } |
6168 | |
6169 | commit_edge_insertions (); |
6170 | |
6171 | /* Look for basic blocks within the prologue insns. */ |
6172 | if (split_prologue_insn |
6173 | && BLOCK_FOR_INSN (insn: split_prologue_insn) == NULL) |
6174 | split_prologue_insn = NULL; |
6175 | if (prologue_insn |
6176 | && BLOCK_FOR_INSN (insn: prologue_insn) == NULL) |
6177 | prologue_insn = NULL; |
6178 | if (split_prologue_insn || prologue_insn) |
6179 | { |
6180 | auto_sbitmap blocks (last_basic_block_for_fn (cfun)); |
6181 | bitmap_clear (blocks); |
6182 | if (split_prologue_insn) |
6183 | bitmap_set_bit (map: blocks, |
6184 | bitno: BLOCK_FOR_INSN (insn: split_prologue_insn)->index); |
6185 | if (prologue_insn) |
6186 | bitmap_set_bit (map: blocks, bitno: BLOCK_FOR_INSN (insn: prologue_insn)->index); |
6187 | find_many_sub_basic_blocks (blocks); |
6188 | } |
6189 | } |
6190 | |
6191 | default_rtl_profile (); |
6192 | |
6193 | /* Emit sibling epilogues before any sibling call sites. */ |
6194 | for (ei = ei_start (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds); |
6195 | (e = ei_safe_edge (i: ei)); |
6196 | ei_next (i: &ei)) |
6197 | { |
6198 | /* Skip those already handled, the ones that run without prologue. */ |
6199 | if (e->flags & EDGE_IGNORE) |
6200 | { |
6201 | e->flags &= ~EDGE_IGNORE; |
6202 | continue; |
6203 | } |
6204 | |
6205 | rtx_insn *insn = BB_END (e->src); |
6206 | |
6207 | if (!(CALL_P (insn) && SIBLING_CALL_P (insn))) |
6208 | continue; |
6209 | |
6210 | if (rtx_insn *ep_seq = targetm.gen_sibcall_epilogue ()) |
6211 | { |
6212 | start_sequence (); |
6213 | emit_note (NOTE_INSN_EPILOGUE_BEG); |
6214 | emit_insn (ep_seq); |
6215 | rtx_insn *seq = get_insns (); |
6216 | end_sequence (); |
6217 | |
6218 | /* Retain a map of the epilogue insns. Used in life analysis to |
6219 | avoid getting rid of sibcall epilogue insns. Do this before we |
6220 | actually emit the sequence. */ |
6221 | record_insns (insns: seq, NULL, hashp: &epilogue_insn_hash); |
6222 | set_insn_locations (seq, epilogue_location); |
6223 | |
6224 | emit_insn_before (seq, insn); |
6225 | |
6226 | find_sub_basic_blocks (BLOCK_FOR_INSN (insn)); |
6227 | } |
6228 | } |
6229 | |
6230 | if (epilogue_seq) |
6231 | { |
6232 | rtx_insn *insn, *next; |
6233 | |
6234 | /* Similarly, move any line notes that appear after the epilogue. |
6235 | There is no need, however, to be quite so anal about the existence |
6236 | of such a note. Also possibly move |
6237 | NOTE_INSN_FUNCTION_BEG notes, as those can be relevant for debug |
6238 | info generation. */ |
6239 | for (insn = epilogue_seq; insn; insn = next) |
6240 | { |
6241 | next = NEXT_INSN (insn); |
6242 | if (NOTE_P (insn) |
6243 | && (NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG)) |
6244 | reorder_insns (insn, insn, PREV_INSN (insn: epilogue_seq)); |
6245 | } |
6246 | } |
6247 | |
6248 | /* Threading the prologue and epilogue changes the artificial refs in the |
6249 | entry and exit blocks, and may invalidate DF info for tail calls. */ |
6250 | if (optimize |
6251 | || flag_optimize_sibling_calls |
6252 | || flag_ipa_icf_functions |
6253 | || in_lto_p) |
6254 | df_update_entry_exit_and_calls (); |
6255 | else |
6256 | { |
6257 | df_update_entry_block_defs (); |
6258 | df_update_exit_block_uses (); |
6259 | } |
6260 | } |
6261 | |
6262 | /* Reposition the prologue-end and epilogue-begin notes after |
6263 | instruction scheduling. */ |
6264 | |
6265 | void |
6266 | reposition_prologue_and_epilogue_notes (void) |
6267 | { |
6268 | if (!targetm.have_prologue () |
6269 | && !targetm.have_epilogue () |
6270 | && !targetm.have_sibcall_epilogue ()) |
6271 | return; |
6272 | |
6273 | /* Since the hash table is created on demand, the fact that it is |
6274 | non-null is a signal that it is non-empty. */ |
6275 | if (prologue_insn_hash != NULL) |
6276 | { |
6277 | size_t len = prologue_insn_hash->elements (); |
6278 | rtx_insn *insn, *last = NULL, *note = NULL; |
6279 | |
6280 | /* Scan from the beginning until we reach the last prologue insn. */ |
6281 | /* ??? While we do have the CFG intact, there are two problems: |
6282 | (1) The prologue can contain loops (typically probing the stack), |
6283 | which means that the end of the prologue isn't in the first bb. |
6284 | (2) Sometimes the PROLOGUE_END note gets pushed into the next bb. */ |
6285 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
6286 | { |
6287 | if (NOTE_P (insn)) |
6288 | { |
6289 | if (NOTE_KIND (insn) == NOTE_INSN_PROLOGUE_END) |
6290 | note = insn; |
6291 | } |
6292 | else if (contains (insn, hash: prologue_insn_hash)) |
6293 | { |
6294 | last = insn; |
6295 | if (--len == 0) |
6296 | break; |
6297 | } |
6298 | } |
6299 | |
6300 | if (last) |
6301 | { |
6302 | if (note == NULL) |
6303 | { |
6304 | /* Scan forward looking for the PROLOGUE_END note. It should |
6305 | be right at the beginning of the block, possibly with other |
6306 | insn notes that got moved there. */ |
6307 | for (note = NEXT_INSN (insn: last); ; note = NEXT_INSN (insn: note)) |
6308 | { |
6309 | if (NOTE_P (note) |
6310 | && NOTE_KIND (note) == NOTE_INSN_PROLOGUE_END) |
6311 | break; |
6312 | } |
6313 | } |
6314 | |
6315 | /* Avoid placing note between CODE_LABEL and BASIC_BLOCK note. */ |
6316 | if (LABEL_P (last)) |
6317 | last = NEXT_INSN (insn: last); |
6318 | reorder_insns (note, note, last); |
6319 | } |
6320 | } |
6321 | |
6322 | if (epilogue_insn_hash != NULL) |
6323 | { |
6324 | edge_iterator ei; |
6325 | edge e; |
6326 | |
6327 | FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) |
6328 | { |
6329 | rtx_insn *insn, *first = NULL, *note = NULL; |
6330 | basic_block bb = e->src; |
6331 | |
6332 | /* Scan from the beginning until we reach the first epilogue insn. */ |
6333 | FOR_BB_INSNS (bb, insn) |
6334 | { |
6335 | if (NOTE_P (insn)) |
6336 | { |
6337 | if (NOTE_KIND (insn) == NOTE_INSN_EPILOGUE_BEG) |
6338 | { |
6339 | note = insn; |
6340 | if (first != NULL) |
6341 | break; |
6342 | } |
6343 | } |
6344 | else if (first == NULL && contains (insn, hash: epilogue_insn_hash)) |
6345 | { |
6346 | first = insn; |
6347 | if (note != NULL) |
6348 | break; |
6349 | } |
6350 | } |
6351 | |
6352 | if (note) |
6353 | { |
6354 | /* If the function has a single basic block, and no real |
6355 | epilogue insns (e.g. sibcall with no cleanup), the |
6356 | epilogue note can get scheduled before the prologue |
6357 | note. If we have frame related prologue insns, having |
6358 | them scanned during the epilogue will result in a crash. |
6359 | In this case re-order the epilogue note to just before |
6360 | the last insn in the block. */ |
6361 | if (first == NULL) |
6362 | first = BB_END (bb); |
6363 | |
6364 | if (PREV_INSN (insn: first) != note) |
6365 | reorder_insns (note, note, PREV_INSN (insn: first)); |
6366 | } |
6367 | } |
6368 | } |
6369 | } |
6370 | |
6371 | /* Returns the name of function declared by FNDECL. */ |
6372 | const char * |
6373 | fndecl_name (tree fndecl) |
6374 | { |
6375 | if (fndecl == NULL) |
6376 | return "(nofn)" ; |
6377 | return lang_hooks.decl_printable_name (fndecl, 1); |
6378 | } |
6379 | |
6380 | /* Returns the name of function FN. */ |
6381 | const char * |
6382 | function_name (struct function *fn) |
6383 | { |
6384 | tree fndecl = (fn == NULL) ? NULL : fn->decl; |
6385 | return fndecl_name (fndecl); |
6386 | } |
6387 | |
6388 | /* Returns the name of the current function. */ |
6389 | const char * |
6390 | current_function_name (void) |
6391 | { |
6392 | return function_name (fn: cfun); |
6393 | } |
6394 | |
6395 | |
6396 | static void |
6397 | rest_of_handle_check_leaf_regs (void) |
6398 | { |
6399 | #ifdef LEAF_REGISTERS |
6400 | crtl->uses_only_leaf_regs |
6401 | = optimize > 0 && only_leaf_regs_used () && leaf_function_p (); |
6402 | #endif |
6403 | } |
6404 | |
6405 | /* Insert a TYPE into the used types hash table of CFUN. */ |
6406 | |
6407 | static void |
6408 | used_types_insert_helper (tree type, struct function *func) |
6409 | { |
6410 | if (type != NULL && func != NULL) |
6411 | { |
6412 | if (func->used_types_hash == NULL) |
6413 | func->used_types_hash = hash_set<tree>::create_ggc (n: 37); |
6414 | |
6415 | func->used_types_hash->add (k: type); |
6416 | } |
6417 | } |
6418 | |
6419 | /* Given a type, insert it into the used hash table in cfun. */ |
6420 | void |
6421 | used_types_insert (tree t) |
6422 | { |
6423 | while (POINTER_TYPE_P (t) || TREE_CODE (t) == ARRAY_TYPE) |
6424 | if (TYPE_NAME (t)) |
6425 | break; |
6426 | else |
6427 | t = TREE_TYPE (t); |
6428 | if (TREE_CODE (t) == ERROR_MARK) |
6429 | return; |
6430 | if (TYPE_NAME (t) == NULL_TREE |
6431 | || TYPE_NAME (t) == TYPE_NAME (TYPE_MAIN_VARIANT (t))) |
6432 | t = TYPE_MAIN_VARIANT (t); |
6433 | if (debug_info_level > DINFO_LEVEL_NONE) |
6434 | { |
6435 | if (cfun) |
6436 | used_types_insert_helper (type: t, func: cfun); |
6437 | else |
6438 | { |
6439 | /* So this might be a type referenced by a global variable. |
6440 | Record that type so that we can later decide to emit its |
6441 | debug information. */ |
6442 | vec_safe_push (v&: types_used_by_cur_var_decl, obj: t); |
6443 | } |
6444 | } |
6445 | } |
6446 | |
6447 | /* Helper to Hash a struct types_used_by_vars_entry. */ |
6448 | |
6449 | static hashval_t |
6450 | hash_types_used_by_vars_entry (const struct types_used_by_vars_entry *entry) |
6451 | { |
6452 | gcc_assert (entry && entry->var_decl && entry->type); |
6453 | |
6454 | return iterative_hash_object (entry->type, |
6455 | iterative_hash_object (entry->var_decl, 0)); |
6456 | } |
6457 | |
6458 | /* Hash function of the types_used_by_vars_entry hash table. */ |
6459 | |
6460 | hashval_t |
6461 | used_type_hasher::hash (types_used_by_vars_entry *entry) |
6462 | { |
6463 | return hash_types_used_by_vars_entry (entry); |
6464 | } |
6465 | |
6466 | /*Equality function of the types_used_by_vars_entry hash table. */ |
6467 | |
6468 | bool |
6469 | used_type_hasher::equal (types_used_by_vars_entry *e1, |
6470 | types_used_by_vars_entry *e2) |
6471 | { |
6472 | return (e1->var_decl == e2->var_decl && e1->type == e2->type); |
6473 | } |
6474 | |
6475 | /* Inserts an entry into the types_used_by_vars_hash hash table. */ |
6476 | |
6477 | void |
6478 | types_used_by_var_decl_insert (tree type, tree var_decl) |
6479 | { |
6480 | if (type != NULL && var_decl != NULL) |
6481 | { |
6482 | types_used_by_vars_entry **slot; |
6483 | struct types_used_by_vars_entry e; |
6484 | e.var_decl = var_decl; |
6485 | e.type = type; |
6486 | if (types_used_by_vars_hash == NULL) |
6487 | types_used_by_vars_hash |
6488 | = hash_table<used_type_hasher>::create_ggc (n: 37); |
6489 | |
6490 | slot = types_used_by_vars_hash->find_slot (value: &e, insert: INSERT); |
6491 | if (*slot == NULL) |
6492 | { |
6493 | struct types_used_by_vars_entry *entry; |
6494 | entry = ggc_alloc<types_used_by_vars_entry> (); |
6495 | entry->type = type; |
6496 | entry->var_decl = var_decl; |
6497 | *slot = entry; |
6498 | } |
6499 | } |
6500 | } |
6501 | |
6502 | namespace { |
6503 | |
6504 | const pass_data pass_data_leaf_regs = |
6505 | { |
6506 | .type: RTL_PASS, /* type */ |
6507 | .name: "*leaf_regs" , /* name */ |
6508 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
6509 | .tv_id: TV_NONE, /* tv_id */ |
6510 | .properties_required: 0, /* properties_required */ |
6511 | .properties_provided: 0, /* properties_provided */ |
6512 | .properties_destroyed: 0, /* properties_destroyed */ |
6513 | .todo_flags_start: 0, /* todo_flags_start */ |
6514 | .todo_flags_finish: 0, /* todo_flags_finish */ |
6515 | }; |
6516 | |
6517 | class pass_leaf_regs : public rtl_opt_pass |
6518 | { |
6519 | public: |
6520 | pass_leaf_regs (gcc::context *ctxt) |
6521 | : rtl_opt_pass (pass_data_leaf_regs, ctxt) |
6522 | {} |
6523 | |
6524 | /* opt_pass methods: */ |
6525 | unsigned int execute (function *) final override |
6526 | { |
6527 | rest_of_handle_check_leaf_regs (); |
6528 | return 0; |
6529 | } |
6530 | |
6531 | }; // class pass_leaf_regs |
6532 | |
6533 | } // anon namespace |
6534 | |
6535 | rtl_opt_pass * |
6536 | make_pass_leaf_regs (gcc::context *ctxt) |
6537 | { |
6538 | return new pass_leaf_regs (ctxt); |
6539 | } |
6540 | |
6541 | static void |
6542 | rest_of_handle_thread_prologue_and_epilogue (function *fun) |
6543 | { |
6544 | /* prepare_shrink_wrap is sensitive to the block structure of the control |
6545 | flow graph, so clean it up first. */ |
6546 | if (optimize) |
6547 | cleanup_cfg (0); |
6548 | |
6549 | /* On some machines, the prologue and epilogue code, or parts thereof, |
6550 | can be represented as RTL. Doing so lets us schedule insns between |
6551 | it and the rest of the code and also allows delayed branch |
6552 | scheduling to operate in the epilogue. */ |
6553 | thread_prologue_and_epilogue_insns (); |
6554 | |
6555 | /* Some non-cold blocks may now be only reachable from cold blocks. |
6556 | Fix that up. */ |
6557 | fixup_partitions (); |
6558 | |
6559 | /* After prologue and epilogue generation, the judgement on whether |
6560 | one memory access onto stack frame may trap or not could change, |
6561 | since we get more exact stack information by now. So try to |
6562 | remove any EH edges here, see PR90259. */ |
6563 | if (fun->can_throw_non_call_exceptions) |
6564 | purge_all_dead_edges (); |
6565 | |
6566 | /* Shrink-wrapping can result in unreachable edges in the epilogue, |
6567 | see PR57320. */ |
6568 | cleanup_cfg (optimize ? CLEANUP_EXPENSIVE : 0); |
6569 | |
6570 | /* The stack usage info is finalized during prologue expansion. */ |
6571 | if (flag_stack_usage_info || flag_callgraph_info) |
6572 | output_stack_usage (); |
6573 | } |
6574 | |
6575 | /* Record a final call to CALLEE at LOCATION. */ |
6576 | |
6577 | void |
6578 | record_final_call (tree callee, location_t location) |
6579 | { |
6580 | struct callinfo_callee datum = { .location: location, .decl: callee }; |
6581 | vec_safe_push (v&: cfun->su->callees, obj: datum); |
6582 | } |
6583 | |
6584 | /* Record a dynamic allocation made for DECL_OR_EXP. */ |
6585 | |
6586 | void |
6587 | record_dynamic_alloc (tree decl_or_exp) |
6588 | { |
6589 | struct callinfo_dalloc datum; |
6590 | |
6591 | if (DECL_P (decl_or_exp)) |
6592 | { |
6593 | datum.location = DECL_SOURCE_LOCATION (decl_or_exp); |
6594 | const char *name = lang_hooks.decl_printable_name (decl_or_exp, 2); |
6595 | const char *dot = strrchr (s: name, c: '.'); |
6596 | if (dot) |
6597 | name = dot + 1; |
6598 | datum.name = ggc_strdup (name); |
6599 | } |
6600 | else |
6601 | { |
6602 | datum.location = EXPR_LOCATION (decl_or_exp); |
6603 | datum.name = NULL; |
6604 | } |
6605 | |
6606 | vec_safe_push (v&: cfun->su->dallocs, obj: datum); |
6607 | } |
6608 | |
6609 | namespace { |
6610 | |
6611 | const pass_data pass_data_thread_prologue_and_epilogue = |
6612 | { |
6613 | .type: RTL_PASS, /* type */ |
6614 | .name: "pro_and_epilogue" , /* name */ |
6615 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
6616 | .tv_id: TV_THREAD_PROLOGUE_AND_EPILOGUE, /* tv_id */ |
6617 | .properties_required: 0, /* properties_required */ |
6618 | .properties_provided: 0, /* properties_provided */ |
6619 | .properties_destroyed: 0, /* properties_destroyed */ |
6620 | .todo_flags_start: 0, /* todo_flags_start */ |
6621 | .todo_flags_finish: ( TODO_df_verify | TODO_df_finish ), /* todo_flags_finish */ |
6622 | }; |
6623 | |
6624 | class pass_thread_prologue_and_epilogue : public rtl_opt_pass |
6625 | { |
6626 | public: |
6627 | pass_thread_prologue_and_epilogue (gcc::context *ctxt) |
6628 | : rtl_opt_pass (pass_data_thread_prologue_and_epilogue, ctxt) |
6629 | {} |
6630 | |
6631 | /* opt_pass methods: */ |
6632 | unsigned int execute (function * fun) final override |
6633 | { |
6634 | rest_of_handle_thread_prologue_and_epilogue (fun); |
6635 | return 0; |
6636 | } |
6637 | |
6638 | }; // class pass_thread_prologue_and_epilogue |
6639 | |
6640 | } // anon namespace |
6641 | |
6642 | rtl_opt_pass * |
6643 | make_pass_thread_prologue_and_epilogue (gcc::context *ctxt) |
6644 | { |
6645 | return new pass_thread_prologue_and_epilogue (ctxt); |
6646 | } |
6647 | |
6648 | namespace { |
6649 | |
6650 | const pass_data pass_data_zero_call_used_regs = |
6651 | { |
6652 | .type: RTL_PASS, /* type */ |
6653 | .name: "zero_call_used_regs" , /* name */ |
6654 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
6655 | .tv_id: TV_NONE, /* tv_id */ |
6656 | .properties_required: 0, /* properties_required */ |
6657 | .properties_provided: 0, /* properties_provided */ |
6658 | .properties_destroyed: 0, /* properties_destroyed */ |
6659 | .todo_flags_start: 0, /* todo_flags_start */ |
6660 | .todo_flags_finish: 0, /* todo_flags_finish */ |
6661 | }; |
6662 | |
6663 | class pass_zero_call_used_regs: public rtl_opt_pass |
6664 | { |
6665 | public: |
6666 | pass_zero_call_used_regs (gcc::context *ctxt) |
6667 | : rtl_opt_pass (pass_data_zero_call_used_regs, ctxt) |
6668 | {} |
6669 | |
6670 | /* opt_pass methods: */ |
6671 | unsigned int execute (function *) final override; |
6672 | |
6673 | }; // class pass_zero_call_used_regs |
6674 | |
6675 | unsigned int |
6676 | pass_zero_call_used_regs::execute (function *fun) |
6677 | { |
6678 | using namespace zero_regs_flags; |
6679 | unsigned int zero_regs_type = UNSET; |
6680 | |
6681 | tree attr_zero_regs = lookup_attribute (attr_name: "zero_call_used_regs" , |
6682 | DECL_ATTRIBUTES (fun->decl)); |
6683 | |
6684 | /* Get the type of zero_call_used_regs from function attribute. |
6685 | We have filtered out invalid attribute values already at this point. */ |
6686 | if (attr_zero_regs) |
6687 | { |
6688 | /* The TREE_VALUE of an attribute is a TREE_LIST whose TREE_VALUE |
6689 | is the attribute argument's value. */ |
6690 | attr_zero_regs = TREE_VALUE (attr_zero_regs); |
6691 | gcc_assert (TREE_CODE (attr_zero_regs) == TREE_LIST); |
6692 | attr_zero_regs = TREE_VALUE (attr_zero_regs); |
6693 | gcc_assert (TREE_CODE (attr_zero_regs) == STRING_CST); |
6694 | |
6695 | for (unsigned int i = 0; zero_call_used_regs_opts[i].name != NULL; ++i) |
6696 | if (strcmp (TREE_STRING_POINTER (attr_zero_regs), |
6697 | s2: zero_call_used_regs_opts[i].name) == 0) |
6698 | { |
6699 | zero_regs_type = zero_call_used_regs_opts[i].flag; |
6700 | break; |
6701 | } |
6702 | } |
6703 | |
6704 | if (!zero_regs_type) |
6705 | zero_regs_type = flag_zero_call_used_regs; |
6706 | |
6707 | /* No need to zero call-used-regs when no user request is present. */ |
6708 | if (!(zero_regs_type & ENABLED)) |
6709 | return 0; |
6710 | |
6711 | edge_iterator ei; |
6712 | edge e; |
6713 | |
6714 | /* This pass needs data flow information. */ |
6715 | df_analyze (); |
6716 | |
6717 | /* Iterate over the function's return instructions and insert any |
6718 | register zeroing required by the -fzero-call-used-regs command-line |
6719 | option or the "zero_call_used_regs" function attribute. */ |
6720 | FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) |
6721 | { |
6722 | rtx_insn *insn = BB_END (e->src); |
6723 | if (JUMP_P (insn) && ANY_RETURN_P (JUMP_LABEL (insn))) |
6724 | gen_call_used_regs_seq (ret: insn, zero_regs_type); |
6725 | } |
6726 | |
6727 | return 0; |
6728 | } |
6729 | |
6730 | } // anon namespace |
6731 | |
6732 | rtl_opt_pass * |
6733 | make_pass_zero_call_used_regs (gcc::context *ctxt) |
6734 | { |
6735 | return new pass_zero_call_used_regs (ctxt); |
6736 | } |
6737 | |
6738 | /* If CONSTRAINT is a matching constraint, then return its number. |
6739 | Otherwise, return -1. */ |
6740 | |
6741 | static int |
6742 | matching_constraint_num (const char *constraint) |
6743 | { |
6744 | if (*constraint == '%') |
6745 | constraint++; |
6746 | |
6747 | if (IN_RANGE (*constraint, '0', '9')) |
6748 | return strtoul (nptr: constraint, NULL, base: 10); |
6749 | |
6750 | return -1; |
6751 | } |
6752 | |
6753 | /* This mini-pass fixes fall-out from SSA in asm statements that have |
6754 | in-out constraints. Say you start with |
6755 | |
6756 | orig = inout; |
6757 | asm ("": "+mr" (inout)); |
6758 | use (orig); |
6759 | |
6760 | which is transformed very early to use explicit output and match operands: |
6761 | |
6762 | orig = inout; |
6763 | asm ("": "=mr" (inout) : "0" (inout)); |
6764 | use (orig); |
6765 | |
6766 | Or, after SSA and copyprop, |
6767 | |
6768 | asm ("": "=mr" (inout_2) : "0" (inout_1)); |
6769 | use (inout_1); |
6770 | |
6771 | Clearly inout_2 and inout_1 can't be coalesced easily anymore, as |
6772 | they represent two separate values, so they will get different pseudo |
6773 | registers during expansion. Then, since the two operands need to match |
6774 | per the constraints, but use different pseudo registers, reload can |
6775 | only register a reload for these operands. But reloads can only be |
6776 | satisfied by hardregs, not by memory, so we need a register for this |
6777 | reload, just because we are presented with non-matching operands. |
6778 | So, even though we allow memory for this operand, no memory can be |
6779 | used for it, just because the two operands don't match. This can |
6780 | cause reload failures on register-starved targets. |
6781 | |
6782 | So it's a symptom of reload not being able to use memory for reloads |
6783 | or, alternatively it's also a symptom of both operands not coming into |
6784 | reload as matching (in which case the pseudo could go to memory just |
6785 | fine, as the alternative allows it, and no reload would be necessary). |
6786 | We fix the latter problem here, by transforming |
6787 | |
6788 | asm ("": "=mr" (inout_2) : "0" (inout_1)); |
6789 | |
6790 | back to |
6791 | |
6792 | inout_2 = inout_1; |
6793 | asm ("": "=mr" (inout_2) : "0" (inout_2)); */ |
6794 | |
6795 | static void |
6796 | match_asm_constraints_1 (rtx_insn *insn, rtx *p_sets, int noutputs) |
6797 | { |
6798 | int i; |
6799 | bool changed = false; |
6800 | rtx op = SET_SRC (p_sets[0]); |
6801 | int ninputs = ASM_OPERANDS_INPUT_LENGTH (op); |
6802 | rtvec inputs = ASM_OPERANDS_INPUT_VEC (op); |
6803 | bool *output_matched = XALLOCAVEC (bool, noutputs); |
6804 | |
6805 | memset (s: output_matched, c: 0, n: noutputs * sizeof (bool)); |
6806 | for (i = 0; i < ninputs; i++) |
6807 | { |
6808 | rtx input, output; |
6809 | rtx_insn *insns; |
6810 | const char *constraint = ASM_OPERANDS_INPUT_CONSTRAINT (op, i); |
6811 | int match, j; |
6812 | |
6813 | match = matching_constraint_num (constraint); |
6814 | if (match < 0) |
6815 | continue; |
6816 | |
6817 | gcc_assert (match < noutputs); |
6818 | output = SET_DEST (p_sets[match]); |
6819 | input = RTVEC_ELT (inputs, i); |
6820 | /* Only do the transformation for pseudos. */ |
6821 | if (! REG_P (output) |
6822 | || rtx_equal_p (output, input) |
6823 | || !(REG_P (input) || SUBREG_P (input) |
6824 | || MEM_P (input) || CONSTANT_P (input)) |
6825 | || !general_operand (input, GET_MODE (output))) |
6826 | continue; |
6827 | |
6828 | /* We can't do anything if the output is also used as input, |
6829 | as we're going to overwrite it. */ |
6830 | for (j = 0; j < ninputs; j++) |
6831 | if (reg_overlap_mentioned_p (output, RTVEC_ELT (inputs, j))) |
6832 | break; |
6833 | if (j != ninputs) |
6834 | continue; |
6835 | |
6836 | /* Avoid changing the same input several times. For |
6837 | asm ("" : "=mr" (out1), "=mr" (out2) : "0" (in), "1" (in)); |
6838 | only change it once (to out1), rather than changing it |
6839 | first to out1 and afterwards to out2. */ |
6840 | if (i > 0) |
6841 | { |
6842 | for (j = 0; j < noutputs; j++) |
6843 | if (output_matched[j] && input == SET_DEST (p_sets[j])) |
6844 | break; |
6845 | if (j != noutputs) |
6846 | continue; |
6847 | } |
6848 | output_matched[match] = true; |
6849 | |
6850 | start_sequence (); |
6851 | emit_move_insn (output, copy_rtx (input)); |
6852 | insns = get_insns (); |
6853 | end_sequence (); |
6854 | emit_insn_before (insns, insn); |
6855 | |
6856 | constraint = ASM_OPERANDS_OUTPUT_CONSTRAINT(SET_SRC(p_sets[match])); |
6857 | bool early_clobber_p = strchr (s: constraint, c: '&') != NULL; |
6858 | |
6859 | /* Now replace all mentions of the input with output. We can't |
6860 | just replace the occurrence in inputs[i], as the register might |
6861 | also be used in some other input (or even in an address of an |
6862 | output), which would mean possibly increasing the number of |
6863 | inputs by one (namely 'output' in addition), which might pose |
6864 | a too complicated problem for reload to solve. E.g. this situation: |
6865 | |
6866 | asm ("" : "=r" (output), "=m" (input) : "0" (input)) |
6867 | |
6868 | Here 'input' is used in two occurrences as input (once for the |
6869 | input operand, once for the address in the second output operand). |
6870 | If we would replace only the occurrence of the input operand (to |
6871 | make the matching) we would be left with this: |
6872 | |
6873 | output = input |
6874 | asm ("" : "=r" (output), "=m" (input) : "0" (output)) |
6875 | |
6876 | Now we suddenly have two different input values (containing the same |
6877 | value, but different pseudos) where we formerly had only one. |
6878 | With more complicated asms this might lead to reload failures |
6879 | which wouldn't have happen without this pass. So, iterate over |
6880 | all operands and replace all occurrences of the register used. |
6881 | |
6882 | However, if one or more of the 'input' uses have a non-matching |
6883 | constraint and the matched output operand is an early clobber |
6884 | operand, then do not replace the input operand, since by definition |
6885 | it conflicts with the output operand and cannot share the same |
6886 | register. See PR89313 for details. */ |
6887 | |
6888 | for (j = 0; j < noutputs; j++) |
6889 | if (!rtx_equal_p (SET_DEST (p_sets[j]), input) |
6890 | && reg_overlap_mentioned_p (input, SET_DEST (p_sets[j]))) |
6891 | SET_DEST (p_sets[j]) = replace_rtx (SET_DEST (p_sets[j]), |
6892 | input, output); |
6893 | for (j = 0; j < ninputs; j++) |
6894 | if (reg_overlap_mentioned_p (input, RTVEC_ELT (inputs, j))) |
6895 | { |
6896 | if (!early_clobber_p |
6897 | || match == matching_constraint_num |
6898 | (ASM_OPERANDS_INPUT_CONSTRAINT (op, j))) |
6899 | RTVEC_ELT (inputs, j) = replace_rtx (RTVEC_ELT (inputs, j), |
6900 | input, output); |
6901 | } |
6902 | |
6903 | changed = true; |
6904 | } |
6905 | |
6906 | if (changed) |
6907 | df_insn_rescan (insn); |
6908 | } |
6909 | |
6910 | /* Add the decl D to the local_decls list of FUN. */ |
6911 | |
6912 | void |
6913 | add_local_decl (struct function *fun, tree d) |
6914 | { |
6915 | gcc_assert (VAR_P (d)); |
6916 | vec_safe_push (v&: fun->local_decls, obj: d); |
6917 | } |
6918 | |
6919 | namespace { |
6920 | |
6921 | const pass_data pass_data_match_asm_constraints = |
6922 | { |
6923 | .type: RTL_PASS, /* type */ |
6924 | .name: "asmcons" , /* name */ |
6925 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
6926 | .tv_id: TV_NONE, /* tv_id */ |
6927 | .properties_required: 0, /* properties_required */ |
6928 | .properties_provided: 0, /* properties_provided */ |
6929 | .properties_destroyed: 0, /* properties_destroyed */ |
6930 | .todo_flags_start: 0, /* todo_flags_start */ |
6931 | .todo_flags_finish: 0, /* todo_flags_finish */ |
6932 | }; |
6933 | |
6934 | class pass_match_asm_constraints : public rtl_opt_pass |
6935 | { |
6936 | public: |
6937 | pass_match_asm_constraints (gcc::context *ctxt) |
6938 | : rtl_opt_pass (pass_data_match_asm_constraints, ctxt) |
6939 | {} |
6940 | |
6941 | /* opt_pass methods: */ |
6942 | unsigned int execute (function *) final override; |
6943 | |
6944 | }; // class pass_match_asm_constraints |
6945 | |
6946 | unsigned |
6947 | pass_match_asm_constraints::execute (function *fun) |
6948 | { |
6949 | basic_block bb; |
6950 | rtx_insn *insn; |
6951 | rtx pat, *p_sets; |
6952 | int noutputs; |
6953 | |
6954 | if (!crtl->has_asm_statement) |
6955 | return 0; |
6956 | |
6957 | df_set_flags (DF_DEFER_INSN_RESCAN); |
6958 | FOR_EACH_BB_FN (bb, fun) |
6959 | { |
6960 | FOR_BB_INSNS (bb, insn) |
6961 | { |
6962 | if (!INSN_P (insn)) |
6963 | continue; |
6964 | |
6965 | pat = PATTERN (insn); |
6966 | if (GET_CODE (pat) == PARALLEL) |
6967 | p_sets = &XVECEXP (pat, 0, 0), noutputs = XVECLEN (pat, 0); |
6968 | else if (GET_CODE (pat) == SET) |
6969 | p_sets = &PATTERN (insn), noutputs = 1; |
6970 | else |
6971 | continue; |
6972 | |
6973 | if (GET_CODE (*p_sets) == SET |
6974 | && GET_CODE (SET_SRC (*p_sets)) == ASM_OPERANDS) |
6975 | match_asm_constraints_1 (insn, p_sets, noutputs); |
6976 | } |
6977 | } |
6978 | |
6979 | return TODO_df_finish; |
6980 | } |
6981 | |
6982 | } // anon namespace |
6983 | |
6984 | rtl_opt_pass * |
6985 | make_pass_match_asm_constraints (gcc::context *ctxt) |
6986 | { |
6987 | return new pass_match_asm_constraints (ctxt); |
6988 | } |
6989 | |
6990 | |
6991 | #include "gt-function.h" |
6992 | |