1 | /* A pass for lowering trees to RTL. |
2 | Copyright (C) 2004-2024 Free Software Foundation, Inc. |
3 | |
4 | This file is part of GCC. |
5 | |
6 | GCC is free software; you can redistribute it and/or modify |
7 | it under the terms of the GNU General Public License as published by |
8 | the Free Software Foundation; either version 3, or (at your option) |
9 | any later version. |
10 | |
11 | GCC is distributed in the hope that it will be useful, |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
14 | GNU General Public License for more details. |
15 | |
16 | You should have received a copy of the GNU General Public License |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | #include "config.h" |
21 | #include "system.h" |
22 | #include "coretypes.h" |
23 | #include "backend.h" |
24 | #include "target.h" |
25 | #include "rtl.h" |
26 | #include "tree.h" |
27 | #include "gimple.h" |
28 | #include "cfghooks.h" |
29 | #include "tree-pass.h" |
30 | #include "memmodel.h" |
31 | #include "tm_p.h" |
32 | #include "ssa.h" |
33 | #include "optabs.h" |
34 | #include "regs.h" /* For reg_renumber. */ |
35 | #include "emit-rtl.h" |
36 | #include "recog.h" |
37 | #include "cgraph.h" |
38 | #include "diagnostic.h" |
39 | #include "fold-const.h" |
40 | #include "varasm.h" |
41 | #include "stor-layout.h" |
42 | #include "stmt.h" |
43 | #include "print-tree.h" |
44 | #include "cfgrtl.h" |
45 | #include "cfganal.h" |
46 | #include "cfgbuild.h" |
47 | #include "cfgcleanup.h" |
48 | #include "dojump.h" |
49 | #include "explow.h" |
50 | #include "calls.h" |
51 | #include "expr.h" |
52 | #include "internal-fn.h" |
53 | #include "tree-eh.h" |
54 | #include "gimple-iterator.h" |
55 | #include "gimple-expr.h" |
56 | #include "gimple-walk.h" |
57 | #include "tree-cfg.h" |
58 | #include "tree-dfa.h" |
59 | #include "tree-ssa.h" |
60 | #include "except.h" |
61 | #include "gimple-pretty-print.h" |
62 | #include "toplev.h" |
63 | #include "debug.h" |
64 | #include "tree-inline.h" |
65 | #include "value-prof.h" |
66 | #include "tree-ssa-live.h" |
67 | #include "tree-outof-ssa.h" |
68 | #include "cfgloop.h" |
69 | #include "insn-attr.h" /* For INSN_SCHEDULING. */ |
70 | #include "stringpool.h" |
71 | #include "attribs.h" |
72 | #include "asan.h" |
73 | #include "tree-ssa-address.h" |
74 | #include "output.h" |
75 | #include "builtins.h" |
76 | #include "opts.h" |
77 | |
78 | /* Some systems use __main in a way incompatible with its use in gcc, in these |
79 | cases use the macros NAME__MAIN to give a quoted symbol and SYMBOL__MAIN to |
80 | give the same symbol without quotes for an alternative entry point. You |
81 | must define both, or neither. */ |
82 | #ifndef NAME__MAIN |
83 | #define NAME__MAIN "__main" |
84 | #endif |
85 | |
86 | /* This variable holds information helping the rewriting of SSA trees |
87 | into RTL. */ |
88 | struct ssaexpand SA; |
89 | |
90 | /* This variable holds the currently expanded gimple statement for purposes |
91 | of comminucating the profile info to the builtin expanders. */ |
92 | gimple *currently_expanding_gimple_stmt; |
93 | |
94 | static rtx expand_debug_expr (tree); |
95 | |
96 | static bool defer_stack_allocation (tree, bool); |
97 | |
98 | static void record_alignment_for_reg_var (unsigned int); |
99 | |
100 | /* Return an expression tree corresponding to the RHS of GIMPLE |
101 | statement STMT. */ |
102 | |
103 | tree |
104 | gimple_assign_rhs_to_tree (gimple *stmt) |
105 | { |
106 | tree t; |
107 | switch (gimple_assign_rhs_class (gs: stmt)) |
108 | { |
109 | case GIMPLE_TERNARY_RHS: |
110 | t = build3 (gimple_assign_rhs_code (gs: stmt), |
111 | TREE_TYPE (gimple_assign_lhs (stmt)), |
112 | gimple_assign_rhs1 (gs: stmt), gimple_assign_rhs2 (gs: stmt), |
113 | gimple_assign_rhs3 (gs: stmt)); |
114 | break; |
115 | case GIMPLE_BINARY_RHS: |
116 | t = build2 (gimple_assign_rhs_code (gs: stmt), |
117 | TREE_TYPE (gimple_assign_lhs (stmt)), |
118 | gimple_assign_rhs1 (gs: stmt), gimple_assign_rhs2 (gs: stmt)); |
119 | break; |
120 | case GIMPLE_UNARY_RHS: |
121 | t = build1 (gimple_assign_rhs_code (gs: stmt), |
122 | TREE_TYPE (gimple_assign_lhs (stmt)), |
123 | gimple_assign_rhs1 (gs: stmt)); |
124 | break; |
125 | case GIMPLE_SINGLE_RHS: |
126 | { |
127 | t = gimple_assign_rhs1 (gs: stmt); |
128 | /* Avoid modifying this tree in place below. */ |
129 | if ((gimple_has_location (g: stmt) && CAN_HAVE_LOCATION_P (t) |
130 | && gimple_location (g: stmt) != EXPR_LOCATION (t)) |
131 | || (gimple_block (g: stmt) && currently_expanding_to_rtl |
132 | && EXPR_P (t))) |
133 | t = copy_node (t); |
134 | break; |
135 | } |
136 | default: |
137 | gcc_unreachable (); |
138 | } |
139 | |
140 | if (gimple_has_location (g: stmt) && CAN_HAVE_LOCATION_P (t)) |
141 | SET_EXPR_LOCATION (t, gimple_location (stmt)); |
142 | |
143 | return t; |
144 | } |
145 | |
146 | |
147 | #ifndef STACK_ALIGNMENT_NEEDED |
148 | #define STACK_ALIGNMENT_NEEDED 1 |
149 | #endif |
150 | |
151 | #define SSAVAR(x) (TREE_CODE (x) == SSA_NAME ? SSA_NAME_VAR (x) : x) |
152 | |
153 | /* Choose either CUR or NEXT as the leader DECL for a partition. |
154 | Prefer ignored decls, to simplify debug dumps and reduce ambiguity |
155 | out of the same user variable being in multiple partitions (this is |
156 | less likely for compiler-introduced temps). */ |
157 | |
158 | static tree |
159 | leader_merge (tree cur, tree next) |
160 | { |
161 | if (cur == NULL || cur == next) |
162 | return next; |
163 | |
164 | if (DECL_P (cur) && DECL_IGNORED_P (cur)) |
165 | return cur; |
166 | |
167 | if (DECL_P (next) && DECL_IGNORED_P (next)) |
168 | return next; |
169 | |
170 | return cur; |
171 | } |
172 | |
173 | /* Associate declaration T with storage space X. If T is no |
174 | SSA name this is exactly SET_DECL_RTL, otherwise make the |
175 | partition of T associated with X. */ |
176 | static inline void |
177 | set_rtl (tree t, rtx x) |
178 | { |
179 | gcc_checking_assert (!x |
180 | || !(TREE_CODE (t) == SSA_NAME || is_gimple_reg (t)) |
181 | || (use_register_for_decl (t) |
182 | ? (REG_P (x) |
183 | || (GET_CODE (x) == CONCAT |
184 | && (REG_P (XEXP (x, 0)) |
185 | || SUBREG_P (XEXP (x, 0))) |
186 | && (REG_P (XEXP (x, 1)) |
187 | || SUBREG_P (XEXP (x, 1)))) |
188 | /* We need to accept PARALLELs for RESUT_DECLs |
189 | because of vector types with BLKmode returned |
190 | in multiple registers, but they are supposed |
191 | to be uncoalesced. */ |
192 | || (GET_CODE (x) == PARALLEL |
193 | && SSAVAR (t) |
194 | && TREE_CODE (SSAVAR (t)) == RESULT_DECL |
195 | && (GET_MODE (x) == BLKmode |
196 | || !flag_tree_coalesce_vars))) |
197 | : (MEM_P (x) || x == pc_rtx |
198 | || (GET_CODE (x) == CONCAT |
199 | && MEM_P (XEXP (x, 0)) |
200 | && MEM_P (XEXP (x, 1)))))); |
201 | /* Check that the RTL for SSA_NAMEs and gimple-reg PARM_DECLs and |
202 | RESULT_DECLs has the expected mode. For memory, we accept |
203 | unpromoted modes, since that's what we're likely to get. For |
204 | PARM_DECLs and RESULT_DECLs, we'll have been called by |
205 | set_parm_rtl, which will give us the default def, so we don't |
206 | have to compute it ourselves. For RESULT_DECLs, we accept mode |
207 | mismatches too, as long as we have BLKmode or are not coalescing |
208 | across variables, so that we don't reject BLKmode PARALLELs or |
209 | unpromoted REGs. */ |
210 | gcc_checking_assert (!x || x == pc_rtx || TREE_CODE (t) != SSA_NAME |
211 | || (SSAVAR (t) |
212 | && TREE_CODE (SSAVAR (t)) == RESULT_DECL |
213 | && (promote_ssa_mode (t, NULL) == BLKmode |
214 | || !flag_tree_coalesce_vars)) |
215 | || !use_register_for_decl (t) |
216 | || GET_MODE (x) == promote_ssa_mode (t, NULL)); |
217 | |
218 | if (x) |
219 | { |
220 | bool skip = false; |
221 | tree cur = NULL_TREE; |
222 | rtx xm = x; |
223 | |
224 | retry: |
225 | if (MEM_P (xm)) |
226 | cur = MEM_EXPR (xm); |
227 | else if (REG_P (xm)) |
228 | cur = REG_EXPR (xm); |
229 | else if (SUBREG_P (xm)) |
230 | { |
231 | gcc_assert (subreg_lowpart_p (xm)); |
232 | xm = SUBREG_REG (xm); |
233 | goto retry; |
234 | } |
235 | else if (GET_CODE (xm) == CONCAT) |
236 | { |
237 | xm = XEXP (xm, 0); |
238 | goto retry; |
239 | } |
240 | else if (GET_CODE (xm) == PARALLEL) |
241 | { |
242 | xm = XVECEXP (xm, 0, 0); |
243 | gcc_assert (GET_CODE (xm) == EXPR_LIST); |
244 | xm = XEXP (xm, 0); |
245 | goto retry; |
246 | } |
247 | else if (xm == pc_rtx) |
248 | skip = true; |
249 | else |
250 | gcc_unreachable (); |
251 | |
252 | tree next = skip ? cur : leader_merge (cur, SSAVAR (t) ? SSAVAR (t) : t); |
253 | |
254 | if (cur != next) |
255 | { |
256 | if (MEM_P (x)) |
257 | set_mem_attributes (x, |
258 | next && TREE_CODE (next) == SSA_NAME |
259 | ? TREE_TYPE (next) |
260 | : next, true); |
261 | else |
262 | set_reg_attrs_for_decl_rtl (t: next, x); |
263 | } |
264 | } |
265 | |
266 | if (TREE_CODE (t) == SSA_NAME) |
267 | { |
268 | int part = var_to_partition (map: SA.map, var: t); |
269 | if (part != NO_PARTITION) |
270 | { |
271 | if (SA.partition_to_pseudo[part]) |
272 | gcc_assert (SA.partition_to_pseudo[part] == x); |
273 | else if (x != pc_rtx) |
274 | SA.partition_to_pseudo[part] = x; |
275 | } |
276 | /* For the benefit of debug information at -O0 (where |
277 | vartracking doesn't run) record the place also in the base |
278 | DECL. For PARMs and RESULTs, do so only when setting the |
279 | default def. */ |
280 | if (x && x != pc_rtx && SSA_NAME_VAR (t) |
281 | && (VAR_P (SSA_NAME_VAR (t)) |
282 | || SSA_NAME_IS_DEFAULT_DEF (t))) |
283 | { |
284 | tree var = SSA_NAME_VAR (t); |
285 | /* If we don't yet have something recorded, just record it now. */ |
286 | if (!DECL_RTL_SET_P (var)) |
287 | SET_DECL_RTL (var, x); |
288 | /* If we have it set already to "multiple places" don't |
289 | change this. */ |
290 | else if (DECL_RTL (var) == pc_rtx) |
291 | ; |
292 | /* If we have something recorded and it's not the same place |
293 | as we want to record now, we have multiple partitions for the |
294 | same base variable, with different places. We can't just |
295 | randomly chose one, hence we have to say that we don't know. |
296 | This only happens with optimization, and there var-tracking |
297 | will figure out the right thing. */ |
298 | else if (DECL_RTL (var) != x) |
299 | SET_DECL_RTL (var, pc_rtx); |
300 | } |
301 | } |
302 | else |
303 | SET_DECL_RTL (t, x); |
304 | } |
305 | |
306 | /* This structure holds data relevant to one variable that will be |
307 | placed in a stack slot. */ |
308 | class stack_var |
309 | { |
310 | public: |
311 | /* The Variable. */ |
312 | tree decl; |
313 | |
314 | /* Initially, the size of the variable. Later, the size of the partition, |
315 | if this variable becomes it's partition's representative. */ |
316 | poly_uint64 size; |
317 | |
318 | /* The *byte* alignment required for this variable. Or as, with the |
319 | size, the alignment for this partition. */ |
320 | unsigned int alignb; |
321 | |
322 | /* The partition representative. */ |
323 | size_t representative; |
324 | |
325 | /* The next stack variable in the partition, or EOC. */ |
326 | size_t next; |
327 | |
328 | /* The numbers of conflicting stack variables. */ |
329 | bitmap conflicts; |
330 | }; |
331 | |
332 | #define EOC ((size_t)-1) |
333 | |
334 | /* We have an array of such objects while deciding allocation. */ |
335 | static class stack_var *stack_vars; |
336 | static size_t stack_vars_alloc; |
337 | static size_t stack_vars_num; |
338 | static hash_map<tree, size_t> *decl_to_stack_part; |
339 | |
340 | /* Conflict bitmaps go on this obstack. This allows us to destroy |
341 | all of them in one big sweep. */ |
342 | static bitmap_obstack stack_var_bitmap_obstack; |
343 | |
344 | /* An array of indices such that stack_vars[stack_vars_sorted[i]].size |
345 | is non-decreasing. */ |
346 | static size_t *stack_vars_sorted; |
347 | |
348 | /* The phase of the stack frame. This is the known misalignment of |
349 | virtual_stack_vars_rtx from PREFERRED_STACK_BOUNDARY. That is, |
350 | (frame_offset+frame_phase) % PREFERRED_STACK_BOUNDARY == 0. */ |
351 | static int frame_phase; |
352 | |
353 | /* Used during expand_used_vars to remember if we saw any decls for |
354 | which we'd like to enable stack smashing protection. */ |
355 | static bool has_protected_decls; |
356 | |
357 | /* Used during expand_used_vars. Remember if we say a character buffer |
358 | smaller than our cutoff threshold. Used for -Wstack-protector. */ |
359 | static bool has_short_buffer; |
360 | |
361 | /* Compute the byte alignment to use for DECL. Ignore alignment |
362 | we can't do with expected alignment of the stack boundary. */ |
363 | |
364 | static unsigned int |
365 | align_local_variable (tree decl, bool really_expand) |
366 | { |
367 | unsigned int align; |
368 | |
369 | if (TREE_CODE (decl) == SSA_NAME) |
370 | { |
371 | tree type = TREE_TYPE (decl); |
372 | machine_mode mode = TYPE_MODE (type); |
373 | |
374 | align = TYPE_ALIGN (type); |
375 | if (mode != BLKmode |
376 | && align < GET_MODE_ALIGNMENT (mode)) |
377 | align = GET_MODE_ALIGNMENT (mode); |
378 | } |
379 | else |
380 | align = LOCAL_DECL_ALIGNMENT (decl); |
381 | |
382 | if (hwasan_sanitize_stack_p ()) |
383 | align = MAX (align, (unsigned) HWASAN_TAG_GRANULE_SIZE * BITS_PER_UNIT); |
384 | |
385 | if (TREE_CODE (decl) != SSA_NAME && really_expand) |
386 | /* Don't change DECL_ALIGN when called from estimated_stack_frame_size. |
387 | That is done before IPA and could bump alignment based on host |
388 | backend even for offloaded code which wants different |
389 | LOCAL_DECL_ALIGNMENT. */ |
390 | SET_DECL_ALIGN (decl, align); |
391 | |
392 | return align / BITS_PER_UNIT; |
393 | } |
394 | |
395 | /* Align given offset BASE with ALIGN. Truncate up if ALIGN_UP is true, |
396 | down otherwise. Return truncated BASE value. */ |
397 | |
398 | static inline unsigned HOST_WIDE_INT |
399 | align_base (HOST_WIDE_INT base, unsigned HOST_WIDE_INT align, bool align_up) |
400 | { |
401 | return align_up ? (base + align - 1) & -align : base & -align; |
402 | } |
403 | |
404 | /* Allocate SIZE bytes at byte alignment ALIGN from the stack frame. |
405 | Return the frame offset. */ |
406 | |
407 | static poly_int64 |
408 | alloc_stack_frame_space (poly_int64 size, unsigned HOST_WIDE_INT align) |
409 | { |
410 | poly_int64 offset, new_frame_offset; |
411 | |
412 | if (FRAME_GROWS_DOWNWARD) |
413 | { |
414 | new_frame_offset |
415 | = aligned_lower_bound (frame_offset - frame_phase - size, |
416 | align) + frame_phase; |
417 | offset = new_frame_offset; |
418 | } |
419 | else |
420 | { |
421 | new_frame_offset |
422 | = aligned_upper_bound (frame_offset - frame_phase, |
423 | align) + frame_phase; |
424 | offset = new_frame_offset; |
425 | new_frame_offset += size; |
426 | } |
427 | frame_offset = new_frame_offset; |
428 | |
429 | if (frame_offset_overflow (frame_offset, cfun->decl)) |
430 | frame_offset = offset = 0; |
431 | |
432 | return offset; |
433 | } |
434 | |
435 | /* Ensure that the stack is aligned to ALIGN bytes. |
436 | Return the new frame offset. */ |
437 | static poly_int64 |
438 | align_frame_offset (unsigned HOST_WIDE_INT align) |
439 | { |
440 | return alloc_stack_frame_space (size: 0, align); |
441 | } |
442 | |
443 | /* Accumulate DECL into STACK_VARS. */ |
444 | |
445 | static void |
446 | add_stack_var (tree decl, bool really_expand) |
447 | { |
448 | class stack_var *v; |
449 | |
450 | if (stack_vars_num >= stack_vars_alloc) |
451 | { |
452 | if (stack_vars_alloc) |
453 | stack_vars_alloc = stack_vars_alloc * 3 / 2; |
454 | else |
455 | stack_vars_alloc = 32; |
456 | stack_vars |
457 | = XRESIZEVEC (class stack_var, stack_vars, stack_vars_alloc); |
458 | } |
459 | if (!decl_to_stack_part) |
460 | decl_to_stack_part = new hash_map<tree, size_t>; |
461 | |
462 | v = &stack_vars[stack_vars_num]; |
463 | decl_to_stack_part->put (k: decl, v: stack_vars_num); |
464 | |
465 | v->decl = decl; |
466 | tree size = TREE_CODE (decl) == SSA_NAME |
467 | ? TYPE_SIZE_UNIT (TREE_TYPE (decl)) |
468 | : DECL_SIZE_UNIT (decl); |
469 | v->size = tree_to_poly_uint64 (size); |
470 | /* Ensure that all variables have size, so that &a != &b for any two |
471 | variables that are simultaneously live. */ |
472 | if (known_eq (v->size, 0U)) |
473 | v->size = 1; |
474 | v->alignb = align_local_variable (decl, really_expand); |
475 | /* An alignment of zero can mightily confuse us later. */ |
476 | gcc_assert (v->alignb != 0); |
477 | |
478 | /* All variables are initially in their own partition. */ |
479 | v->representative = stack_vars_num; |
480 | v->next = EOC; |
481 | |
482 | /* All variables initially conflict with no other. */ |
483 | v->conflicts = NULL; |
484 | |
485 | /* Ensure that this decl doesn't get put onto the list twice. */ |
486 | set_rtl (t: decl, x: pc_rtx); |
487 | |
488 | stack_vars_num++; |
489 | } |
490 | |
491 | /* Make the decls associated with luid's X and Y conflict. */ |
492 | |
493 | static void |
494 | add_stack_var_conflict (size_t x, size_t y) |
495 | { |
496 | class stack_var *a = &stack_vars[x]; |
497 | class stack_var *b = &stack_vars[y]; |
498 | if (x == y) |
499 | return; |
500 | if (!a->conflicts) |
501 | a->conflicts = BITMAP_ALLOC (obstack: &stack_var_bitmap_obstack); |
502 | if (!b->conflicts) |
503 | b->conflicts = BITMAP_ALLOC (obstack: &stack_var_bitmap_obstack); |
504 | bitmap_set_bit (a->conflicts, y); |
505 | bitmap_set_bit (b->conflicts, x); |
506 | } |
507 | |
508 | /* Check whether the decls associated with luid's X and Y conflict. */ |
509 | |
510 | static bool |
511 | stack_var_conflict_p (size_t x, size_t y) |
512 | { |
513 | class stack_var *a = &stack_vars[x]; |
514 | class stack_var *b = &stack_vars[y]; |
515 | if (x == y) |
516 | return false; |
517 | /* Partitions containing an SSA name result from gimple registers |
518 | with things like unsupported modes. They are top-level and |
519 | hence conflict with everything else. */ |
520 | if (TREE_CODE (a->decl) == SSA_NAME || TREE_CODE (b->decl) == SSA_NAME) |
521 | return true; |
522 | |
523 | if (!a->conflicts || !b->conflicts) |
524 | return false; |
525 | return bitmap_bit_p (a->conflicts, y); |
526 | } |
527 | |
528 | /* Callback for walk_stmt_ops. If OP is a decl touched by add_stack_var |
529 | enter its partition number into bitmap DATA. */ |
530 | |
531 | static bool |
532 | visit_op (gimple *, tree op, tree, void *data) |
533 | { |
534 | bitmap active = (bitmap)data; |
535 | op = get_base_address (t: op); |
536 | if (op |
537 | && DECL_P (op) |
538 | && DECL_RTL_IF_SET (op) == pc_rtx) |
539 | { |
540 | size_t *v = decl_to_stack_part->get (k: op); |
541 | if (v) |
542 | bitmap_set_bit (active, *v); |
543 | } |
544 | return false; |
545 | } |
546 | |
547 | /* Callback for walk_stmt_ops. If OP is a decl touched by add_stack_var |
548 | record conflicts between it and all currently active other partitions |
549 | from bitmap DATA. */ |
550 | |
551 | static bool |
552 | visit_conflict (gimple *, tree op, tree, void *data) |
553 | { |
554 | bitmap active = (bitmap)data; |
555 | op = get_base_address (t: op); |
556 | if (op |
557 | && DECL_P (op) |
558 | && DECL_RTL_IF_SET (op) == pc_rtx) |
559 | { |
560 | size_t *v = decl_to_stack_part->get (k: op); |
561 | if (v && bitmap_set_bit (active, *v)) |
562 | { |
563 | size_t num = *v; |
564 | bitmap_iterator bi; |
565 | unsigned i; |
566 | gcc_assert (num < stack_vars_num); |
567 | EXECUTE_IF_SET_IN_BITMAP (active, 0, i, bi) |
568 | add_stack_var_conflict (x: num, y: i); |
569 | } |
570 | } |
571 | return false; |
572 | } |
573 | |
574 | /* Helper function for add_scope_conflicts_1. For USE on |
575 | a stmt, if it is a SSA_NAME and in its SSA_NAME_DEF_STMT is known to be |
576 | based on some ADDR_EXPR, invoke VISIT on that ADDR_EXPR. */ |
577 | |
578 | static inline void |
579 | add_scope_conflicts_2 (tree use, bitmap work, |
580 | walk_stmt_load_store_addr_fn visit) |
581 | { |
582 | if (TREE_CODE (use) == SSA_NAME |
583 | && (POINTER_TYPE_P (TREE_TYPE (use)) |
584 | || INTEGRAL_TYPE_P (TREE_TYPE (use)))) |
585 | { |
586 | gimple *g = SSA_NAME_DEF_STMT (use); |
587 | if (is_gimple_assign (gs: g)) |
588 | if (tree op = gimple_assign_rhs1 (gs: g)) |
589 | if (TREE_CODE (op) == ADDR_EXPR) |
590 | visit (g, TREE_OPERAND (op, 0), op, work); |
591 | } |
592 | } |
593 | |
594 | /* Helper routine for add_scope_conflicts, calculating the active partitions |
595 | at the end of BB, leaving the result in WORK. We're called to generate |
596 | conflicts when FOR_CONFLICT is true, otherwise we're just tracking |
597 | liveness. */ |
598 | |
599 | static void |
600 | add_scope_conflicts_1 (basic_block bb, bitmap work, bool for_conflict) |
601 | { |
602 | edge e; |
603 | edge_iterator ei; |
604 | gimple_stmt_iterator gsi; |
605 | walk_stmt_load_store_addr_fn visit; |
606 | use_operand_p use_p; |
607 | ssa_op_iter iter; |
608 | |
609 | bitmap_clear (work); |
610 | FOR_EACH_EDGE (e, ei, bb->preds) |
611 | bitmap_ior_into (work, (bitmap)e->src->aux); |
612 | |
613 | visit = visit_op; |
614 | |
615 | for (gsi = gsi_start_phis (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
616 | { |
617 | gimple *stmt = gsi_stmt (i: gsi); |
618 | gphi *phi = as_a <gphi *> (p: stmt); |
619 | walk_stmt_load_store_addr_ops (stmt, work, NULL, NULL, visit); |
620 | FOR_EACH_PHI_ARG (use_p, phi, iter, SSA_OP_USE) |
621 | add_scope_conflicts_2 (USE_FROM_PTR (use_p), work, visit); |
622 | } |
623 | for (gsi = gsi_after_labels (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
624 | { |
625 | gimple *stmt = gsi_stmt (i: gsi); |
626 | |
627 | if (gimple_clobber_p (s: stmt)) |
628 | { |
629 | tree lhs = gimple_assign_lhs (gs: stmt); |
630 | size_t *v; |
631 | /* Nested function lowering might introduce LHSs |
632 | that are COMPONENT_REFs. */ |
633 | if (!VAR_P (lhs)) |
634 | continue; |
635 | if (DECL_RTL_IF_SET (lhs) == pc_rtx |
636 | && (v = decl_to_stack_part->get (k: lhs))) |
637 | bitmap_clear_bit (work, *v); |
638 | } |
639 | else if (!is_gimple_debug (gs: stmt)) |
640 | { |
641 | if (for_conflict && visit == visit_op) |
642 | { |
643 | /* If this is the first real instruction in this BB we need |
644 | to add conflicts for everything live at this point now. |
645 | Unlike classical liveness for named objects we can't |
646 | rely on seeing a def/use of the names we're interested in. |
647 | There might merely be indirect loads/stores. We'd not add any |
648 | conflicts for such partitions. */ |
649 | bitmap_iterator bi; |
650 | unsigned i; |
651 | EXECUTE_IF_SET_IN_BITMAP (work, 0, i, bi) |
652 | { |
653 | class stack_var *a = &stack_vars[i]; |
654 | if (!a->conflicts) |
655 | a->conflicts = BITMAP_ALLOC (obstack: &stack_var_bitmap_obstack); |
656 | bitmap_ior_into (a->conflicts, work); |
657 | } |
658 | visit = visit_conflict; |
659 | } |
660 | walk_stmt_load_store_addr_ops (stmt, work, visit, visit, visit); |
661 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) |
662 | add_scope_conflicts_2 (USE_FROM_PTR (use_p), work, visit); |
663 | } |
664 | } |
665 | } |
666 | |
667 | /* Generate stack partition conflicts between all partitions that are |
668 | simultaneously live. */ |
669 | |
670 | static void |
671 | add_scope_conflicts (void) |
672 | { |
673 | basic_block bb; |
674 | bool changed; |
675 | bitmap work = BITMAP_ALLOC (NULL); |
676 | int *rpo; |
677 | int n_bbs; |
678 | |
679 | /* We approximate the live range of a stack variable by taking the first |
680 | mention of its name as starting point(s), and by the end-of-scope |
681 | death clobber added by gimplify as ending point(s) of the range. |
682 | This overapproximates in the case we for instance moved an address-taken |
683 | operation upward, without also moving a dereference to it upwards. |
684 | But it's conservatively correct as a variable never can hold values |
685 | before its name is mentioned at least once. |
686 | |
687 | We then do a mostly classical bitmap liveness algorithm. */ |
688 | |
689 | FOR_ALL_BB_FN (bb, cfun) |
690 | bb->aux = BITMAP_ALLOC (obstack: &stack_var_bitmap_obstack); |
691 | |
692 | rpo = XNEWVEC (int, last_basic_block_for_fn (cfun)); |
693 | n_bbs = pre_and_rev_post_order_compute (NULL, rpo, false); |
694 | |
695 | changed = true; |
696 | while (changed) |
697 | { |
698 | int i; |
699 | changed = false; |
700 | for (i = 0; i < n_bbs; i++) |
701 | { |
702 | bitmap active; |
703 | bb = BASIC_BLOCK_FOR_FN (cfun, rpo[i]); |
704 | active = (bitmap)bb->aux; |
705 | add_scope_conflicts_1 (bb, work, for_conflict: false); |
706 | if (bitmap_ior_into (active, work)) |
707 | changed = true; |
708 | } |
709 | } |
710 | |
711 | FOR_EACH_BB_FN (bb, cfun) |
712 | add_scope_conflicts_1 (bb, work, for_conflict: true); |
713 | |
714 | free (ptr: rpo); |
715 | BITMAP_FREE (work); |
716 | FOR_ALL_BB_FN (bb, cfun) |
717 | BITMAP_FREE (bb->aux); |
718 | } |
719 | |
720 | /* A subroutine of partition_stack_vars. A comparison function for qsort, |
721 | sorting an array of indices by the properties of the object. */ |
722 | |
723 | static int |
724 | stack_var_cmp (const void *a, const void *b) |
725 | { |
726 | size_t ia = *(const size_t *)a; |
727 | size_t ib = *(const size_t *)b; |
728 | unsigned int aligna = stack_vars[ia].alignb; |
729 | unsigned int alignb = stack_vars[ib].alignb; |
730 | poly_int64 sizea = stack_vars[ia].size; |
731 | poly_int64 sizeb = stack_vars[ib].size; |
732 | tree decla = stack_vars[ia].decl; |
733 | tree declb = stack_vars[ib].decl; |
734 | bool largea, largeb; |
735 | unsigned int uida, uidb; |
736 | |
737 | /* Primary compare on "large" alignment. Large comes first. */ |
738 | largea = (aligna * BITS_PER_UNIT > MAX_SUPPORTED_STACK_ALIGNMENT); |
739 | largeb = (alignb * BITS_PER_UNIT > MAX_SUPPORTED_STACK_ALIGNMENT); |
740 | if (largea != largeb) |
741 | return (int)largeb - (int)largea; |
742 | |
743 | /* Secondary compare on size, decreasing */ |
744 | int diff = compare_sizes_for_sort (a: sizeb, b: sizea); |
745 | if (diff != 0) |
746 | return diff; |
747 | |
748 | /* Tertiary compare on true alignment, decreasing. */ |
749 | if (aligna < alignb) |
750 | return -1; |
751 | if (aligna > alignb) |
752 | return 1; |
753 | |
754 | /* Final compare on ID for sort stability, increasing. |
755 | Two SSA names are compared by their version, SSA names come before |
756 | non-SSA names, and two normal decls are compared by their DECL_UID. */ |
757 | if (TREE_CODE (decla) == SSA_NAME) |
758 | { |
759 | if (TREE_CODE (declb) == SSA_NAME) |
760 | uida = SSA_NAME_VERSION (decla), uidb = SSA_NAME_VERSION (declb); |
761 | else |
762 | return -1; |
763 | } |
764 | else if (TREE_CODE (declb) == SSA_NAME) |
765 | return 1; |
766 | else |
767 | uida = DECL_UID (decla), uidb = DECL_UID (declb); |
768 | if (uida < uidb) |
769 | return 1; |
770 | if (uida > uidb) |
771 | return -1; |
772 | return 0; |
773 | } |
774 | |
775 | struct part_traits : unbounded_int_hashmap_traits <size_t, bitmap> {}; |
776 | typedef hash_map<size_t, bitmap, part_traits> part_hashmap; |
777 | |
778 | /* If the points-to solution *PI points to variables that are in a partition |
779 | together with other variables add all partition members to the pointed-to |
780 | variables bitmap. */ |
781 | |
782 | static void |
783 | add_partitioned_vars_to_ptset (struct pt_solution *pt, |
784 | part_hashmap *decls_to_partitions, |
785 | hash_set<bitmap> *visited, bitmap temp) |
786 | { |
787 | bitmap_iterator bi; |
788 | unsigned i; |
789 | bitmap *part; |
790 | |
791 | if (pt->anything |
792 | || pt->vars == NULL |
793 | /* The pointed-to vars bitmap is shared, it is enough to |
794 | visit it once. */ |
795 | || visited->add (k: pt->vars)) |
796 | return; |
797 | |
798 | bitmap_clear (temp); |
799 | |
800 | /* By using a temporary bitmap to store all members of the partitions |
801 | we have to add we make sure to visit each of the partitions only |
802 | once. */ |
803 | EXECUTE_IF_SET_IN_BITMAP (pt->vars, 0, i, bi) |
804 | if ((!temp |
805 | || !bitmap_bit_p (temp, i)) |
806 | && (part = decls_to_partitions->get (k: i))) |
807 | bitmap_ior_into (temp, *part); |
808 | if (!bitmap_empty_p (map: temp)) |
809 | bitmap_ior_into (pt->vars, temp); |
810 | } |
811 | |
812 | /* Update points-to sets based on partition info, so we can use them on RTL. |
813 | The bitmaps representing stack partitions will be saved until expand, |
814 | where partitioned decls used as bases in memory expressions will be |
815 | rewritten. |
816 | |
817 | It is not necessary to update TBAA info on accesses to the coalesced |
818 | storage since our memory model doesn't allow TBAA to be used for |
819 | WAW or WAR dependences. For RAW when the write is to an old object |
820 | the new object would not have been initialized at the point of the |
821 | read, invoking undefined behavior. */ |
822 | |
823 | static void |
824 | update_alias_info_with_stack_vars (void) |
825 | { |
826 | part_hashmap *decls_to_partitions = NULL; |
827 | size_t i, j; |
828 | tree var = NULL_TREE; |
829 | |
830 | for (i = 0; i < stack_vars_num; i++) |
831 | { |
832 | bitmap part = NULL; |
833 | tree name; |
834 | struct ptr_info_def *pi; |
835 | |
836 | /* Not interested in partitions with single variable. */ |
837 | if (stack_vars[i].representative != i |
838 | || stack_vars[i].next == EOC) |
839 | continue; |
840 | |
841 | if (!decls_to_partitions) |
842 | { |
843 | decls_to_partitions = new part_hashmap; |
844 | cfun->gimple_df->decls_to_pointers = new hash_map<tree, tree>; |
845 | } |
846 | |
847 | /* Create an SSA_NAME that points to the partition for use |
848 | as base during alias-oracle queries on RTL for bases that |
849 | have been partitioned. */ |
850 | if (var == NULL_TREE) |
851 | var = create_tmp_var (ptr_type_node); |
852 | name = make_ssa_name (var); |
853 | |
854 | /* Create bitmaps representing partitions. They will be used for |
855 | points-to sets later, so use GGC alloc. */ |
856 | part = BITMAP_GGC_ALLOC (); |
857 | for (j = i; j != EOC; j = stack_vars[j].next) |
858 | { |
859 | tree decl = stack_vars[j].decl; |
860 | unsigned int uid = DECL_PT_UID (decl); |
861 | bitmap_set_bit (part, uid); |
862 | decls_to_partitions->put (k: uid, v: part); |
863 | cfun->gimple_df->decls_to_pointers->put (k: decl, v: name); |
864 | if (TREE_ADDRESSABLE (decl)) |
865 | TREE_ADDRESSABLE (name) = 1; |
866 | } |
867 | |
868 | /* Make the SSA name point to all partition members. */ |
869 | pi = get_ptr_info (name); |
870 | pt_solution_set (&pi->pt, part, false); |
871 | } |
872 | |
873 | /* Make all points-to sets that contain one member of a partition |
874 | contain all members of the partition. */ |
875 | if (decls_to_partitions) |
876 | { |
877 | unsigned i; |
878 | tree name; |
879 | hash_set<bitmap> visited; |
880 | bitmap temp = BITMAP_ALLOC (obstack: &stack_var_bitmap_obstack); |
881 | |
882 | FOR_EACH_SSA_NAME (i, name, cfun) |
883 | { |
884 | struct ptr_info_def *pi; |
885 | |
886 | if (POINTER_TYPE_P (TREE_TYPE (name)) |
887 | && ((pi = SSA_NAME_PTR_INFO (name)) != NULL)) |
888 | add_partitioned_vars_to_ptset (pt: &pi->pt, decls_to_partitions, |
889 | visited: &visited, temp); |
890 | } |
891 | |
892 | add_partitioned_vars_to_ptset (pt: &cfun->gimple_df->escaped, |
893 | decls_to_partitions, visited: &visited, temp); |
894 | add_partitioned_vars_to_ptset (pt: &cfun->gimple_df->escaped_return, |
895 | decls_to_partitions, visited: &visited, temp); |
896 | delete decls_to_partitions; |
897 | BITMAP_FREE (temp); |
898 | } |
899 | } |
900 | |
901 | /* A subroutine of partition_stack_vars. The UNION portion of a UNION/FIND |
902 | partitioning algorithm. Partitions A and B are known to be non-conflicting. |
903 | Merge them into a single partition A. */ |
904 | |
905 | static void |
906 | union_stack_vars (size_t a, size_t b) |
907 | { |
908 | class stack_var *vb = &stack_vars[b]; |
909 | bitmap_iterator bi; |
910 | unsigned u; |
911 | |
912 | gcc_assert (stack_vars[b].next == EOC); |
913 | /* Add B to A's partition. */ |
914 | stack_vars[b].next = stack_vars[a].next; |
915 | stack_vars[b].representative = a; |
916 | stack_vars[a].next = b; |
917 | |
918 | /* Make sure A is big enough to hold B. */ |
919 | stack_vars[a].size = upper_bound (a: stack_vars[a].size, b: stack_vars[b].size); |
920 | |
921 | /* Update the required alignment of partition A to account for B. */ |
922 | if (stack_vars[a].alignb < stack_vars[b].alignb) |
923 | stack_vars[a].alignb = stack_vars[b].alignb; |
924 | |
925 | /* Update the interference graph and merge the conflicts. */ |
926 | if (vb->conflicts) |
927 | { |
928 | EXECUTE_IF_SET_IN_BITMAP (vb->conflicts, 0, u, bi) |
929 | add_stack_var_conflict (x: a, y: stack_vars[u].representative); |
930 | BITMAP_FREE (vb->conflicts); |
931 | } |
932 | } |
933 | |
934 | /* A subroutine of expand_used_vars. Binpack the variables into |
935 | partitions constrained by the interference graph. The overall |
936 | algorithm used is as follows: |
937 | |
938 | Sort the objects by size in descending order. |
939 | For each object A { |
940 | S = size(A) |
941 | O = 0 |
942 | loop { |
943 | Look for the largest non-conflicting object B with size <= S. |
944 | UNION (A, B) |
945 | } |
946 | } |
947 | */ |
948 | |
949 | static void |
950 | partition_stack_vars (void) |
951 | { |
952 | size_t si, sj, n = stack_vars_num; |
953 | |
954 | stack_vars_sorted = XNEWVEC (size_t, stack_vars_num); |
955 | for (si = 0; si < n; ++si) |
956 | stack_vars_sorted[si] = si; |
957 | |
958 | if (n == 1) |
959 | return; |
960 | |
961 | qsort (stack_vars_sorted, n, sizeof (size_t), stack_var_cmp); |
962 | |
963 | for (si = 0; si < n; ++si) |
964 | { |
965 | size_t i = stack_vars_sorted[si]; |
966 | unsigned int ialign = stack_vars[i].alignb; |
967 | poly_int64 isize = stack_vars[i].size; |
968 | |
969 | /* Ignore objects that aren't partition representatives. If we |
970 | see a var that is not a partition representative, it must |
971 | have been merged earlier. */ |
972 | if (stack_vars[i].representative != i) |
973 | continue; |
974 | |
975 | for (sj = si + 1; sj < n; ++sj) |
976 | { |
977 | size_t j = stack_vars_sorted[sj]; |
978 | unsigned int jalign = stack_vars[j].alignb; |
979 | poly_int64 jsize = stack_vars[j].size; |
980 | |
981 | /* Ignore objects that aren't partition representatives. */ |
982 | if (stack_vars[j].representative != j) |
983 | continue; |
984 | |
985 | /* Do not mix objects of "small" (supported) alignment |
986 | and "large" (unsupported) alignment. */ |
987 | if ((ialign * BITS_PER_UNIT <= MAX_SUPPORTED_STACK_ALIGNMENT) |
988 | != (jalign * BITS_PER_UNIT <= MAX_SUPPORTED_STACK_ALIGNMENT)) |
989 | break; |
990 | |
991 | /* For Address Sanitizer do not mix objects with different |
992 | sizes, as the shorter vars wouldn't be adequately protected. |
993 | Don't do that for "large" (unsupported) alignment objects, |
994 | those aren't protected anyway. */ |
995 | if (asan_sanitize_stack_p () |
996 | && maybe_ne (a: isize, b: jsize) |
997 | && ialign * BITS_PER_UNIT <= MAX_SUPPORTED_STACK_ALIGNMENT) |
998 | break; |
999 | |
1000 | /* Ignore conflicting objects. */ |
1001 | if (stack_var_conflict_p (x: i, y: j)) |
1002 | continue; |
1003 | |
1004 | /* UNION the objects, placing J at OFFSET. */ |
1005 | union_stack_vars (a: i, b: j); |
1006 | } |
1007 | } |
1008 | |
1009 | update_alias_info_with_stack_vars (); |
1010 | } |
1011 | |
1012 | /* A debugging aid for expand_used_vars. Dump the generated partitions. */ |
1013 | |
1014 | static void |
1015 | dump_stack_var_partition (void) |
1016 | { |
1017 | size_t si, i, j, n = stack_vars_num; |
1018 | |
1019 | for (si = 0; si < n; ++si) |
1020 | { |
1021 | i = stack_vars_sorted[si]; |
1022 | |
1023 | /* Skip variables that aren't partition representatives, for now. */ |
1024 | if (stack_vars[i].representative != i) |
1025 | continue; |
1026 | |
1027 | fprintf (stream: dump_file, format: "Partition " HOST_SIZE_T_PRINT_UNSIGNED ": size " , |
1028 | (fmt_size_t) i); |
1029 | print_dec (value: stack_vars[i].size, file: dump_file); |
1030 | fprintf (stream: dump_file, format: " align %u\n" , stack_vars[i].alignb); |
1031 | |
1032 | for (j = i; j != EOC; j = stack_vars[j].next) |
1033 | { |
1034 | fputc (c: '\t', stream: dump_file); |
1035 | print_generic_expr (dump_file, stack_vars[j].decl, dump_flags); |
1036 | } |
1037 | fputc (c: '\n', stream: dump_file); |
1038 | } |
1039 | } |
1040 | |
1041 | /* Assign rtl to DECL at BASE + OFFSET. */ |
1042 | |
1043 | static void |
1044 | expand_one_stack_var_at (tree decl, rtx base, unsigned base_align, |
1045 | poly_int64 offset) |
1046 | { |
1047 | unsigned align; |
1048 | rtx x; |
1049 | |
1050 | /* If this fails, we've overflowed the stack frame. Error nicely? */ |
1051 | gcc_assert (known_eq (offset, trunc_int_for_mode (offset, Pmode))); |
1052 | |
1053 | if (hwasan_sanitize_stack_p ()) |
1054 | x = targetm.memtag.add_tag (base, offset, |
1055 | hwasan_current_frame_tag ()); |
1056 | else |
1057 | x = plus_constant (Pmode, base, offset); |
1058 | |
1059 | x = gen_rtx_MEM (TREE_CODE (decl) == SSA_NAME |
1060 | ? TYPE_MODE (TREE_TYPE (decl)) |
1061 | : DECL_MODE (decl), x); |
1062 | |
1063 | /* Set alignment we actually gave this decl if it isn't an SSA name. |
1064 | If it is we generate stack slots only accidentally so it isn't as |
1065 | important, we'll simply set the alignment directly on the MEM. */ |
1066 | |
1067 | if (stack_vars_base_reg_p (base)) |
1068 | offset -= frame_phase; |
1069 | align = known_alignment (a: offset); |
1070 | align *= BITS_PER_UNIT; |
1071 | if (align == 0 || align > base_align) |
1072 | align = base_align; |
1073 | |
1074 | if (TREE_CODE (decl) != SSA_NAME) |
1075 | { |
1076 | /* One would think that we could assert that we're not decreasing |
1077 | alignment here, but (at least) the i386 port does exactly this |
1078 | via the MINIMUM_ALIGNMENT hook. */ |
1079 | |
1080 | SET_DECL_ALIGN (decl, align); |
1081 | DECL_USER_ALIGN (decl) = 0; |
1082 | } |
1083 | |
1084 | set_rtl (t: decl, x); |
1085 | |
1086 | set_mem_align (x, align); |
1087 | } |
1088 | |
1089 | class stack_vars_data |
1090 | { |
1091 | public: |
1092 | /* Vector of offset pairs, always end of some padding followed |
1093 | by start of the padding that needs Address Sanitizer protection. |
1094 | The vector is in reversed, highest offset pairs come first. */ |
1095 | auto_vec<HOST_WIDE_INT> asan_vec; |
1096 | |
1097 | /* Vector of partition representative decls in between the paddings. */ |
1098 | auto_vec<tree> asan_decl_vec; |
1099 | |
1100 | /* Base pseudo register for Address Sanitizer protected automatic vars. */ |
1101 | rtx asan_base; |
1102 | |
1103 | /* Alignment needed for the Address Sanitizer protected automatic vars. */ |
1104 | unsigned int asan_alignb; |
1105 | }; |
1106 | |
1107 | /* A subroutine of expand_used_vars. Give each partition representative |
1108 | a unique location within the stack frame. Update each partition member |
1109 | with that location. */ |
1110 | static void |
1111 | expand_stack_vars (bool (*pred) (size_t), class stack_vars_data *data) |
1112 | { |
1113 | size_t si, i, j, n = stack_vars_num; |
1114 | poly_uint64 large_size = 0, large_alloc = 0; |
1115 | rtx large_base = NULL; |
1116 | rtx large_untagged_base = NULL; |
1117 | unsigned large_align = 0; |
1118 | bool large_allocation_done = false; |
1119 | tree decl; |
1120 | |
1121 | /* Determine if there are any variables requiring "large" alignment. |
1122 | Since these are dynamically allocated, we only process these if |
1123 | no predicate involved. */ |
1124 | large_align = stack_vars[stack_vars_sorted[0]].alignb * BITS_PER_UNIT; |
1125 | if (pred == NULL && large_align > MAX_SUPPORTED_STACK_ALIGNMENT) |
1126 | { |
1127 | /* Find the total size of these variables. */ |
1128 | for (si = 0; si < n; ++si) |
1129 | { |
1130 | unsigned alignb; |
1131 | |
1132 | i = stack_vars_sorted[si]; |
1133 | alignb = stack_vars[i].alignb; |
1134 | |
1135 | /* All "large" alignment decls come before all "small" alignment |
1136 | decls, but "large" alignment decls are not sorted based on |
1137 | their alignment. Increase large_align to track the largest |
1138 | required alignment. */ |
1139 | if ((alignb * BITS_PER_UNIT) > large_align) |
1140 | large_align = alignb * BITS_PER_UNIT; |
1141 | |
1142 | /* Stop when we get to the first decl with "small" alignment. */ |
1143 | if (alignb * BITS_PER_UNIT <= MAX_SUPPORTED_STACK_ALIGNMENT) |
1144 | break; |
1145 | |
1146 | /* Skip variables that aren't partition representatives. */ |
1147 | if (stack_vars[i].representative != i) |
1148 | continue; |
1149 | |
1150 | /* Skip variables that have already had rtl assigned. See also |
1151 | add_stack_var where we perpetrate this pc_rtx hack. */ |
1152 | decl = stack_vars[i].decl; |
1153 | if (TREE_CODE (decl) == SSA_NAME |
1154 | ? SA.partition_to_pseudo[var_to_partition (map: SA.map, var: decl)] != NULL_RTX |
1155 | : DECL_RTL (decl) != pc_rtx) |
1156 | continue; |
1157 | |
1158 | large_size = aligned_upper_bound (value: large_size, align: alignb); |
1159 | large_size += stack_vars[i].size; |
1160 | } |
1161 | } |
1162 | |
1163 | for (si = 0; si < n; ++si) |
1164 | { |
1165 | rtx base; |
1166 | unsigned base_align, alignb; |
1167 | poly_int64 offset = 0; |
1168 | |
1169 | i = stack_vars_sorted[si]; |
1170 | |
1171 | /* Skip variables that aren't partition representatives, for now. */ |
1172 | if (stack_vars[i].representative != i) |
1173 | continue; |
1174 | |
1175 | /* Skip variables that have already had rtl assigned. See also |
1176 | add_stack_var where we perpetrate this pc_rtx hack. */ |
1177 | decl = stack_vars[i].decl; |
1178 | if (TREE_CODE (decl) == SSA_NAME |
1179 | ? SA.partition_to_pseudo[var_to_partition (map: SA.map, var: decl)] != NULL_RTX |
1180 | : DECL_RTL (decl) != pc_rtx) |
1181 | continue; |
1182 | |
1183 | /* Check the predicate to see whether this variable should be |
1184 | allocated in this pass. */ |
1185 | if (pred && !pred (i)) |
1186 | continue; |
1187 | |
1188 | base = (hwasan_sanitize_stack_p () |
1189 | ? hwasan_frame_base () |
1190 | : virtual_stack_vars_rtx); |
1191 | alignb = stack_vars[i].alignb; |
1192 | if (alignb * BITS_PER_UNIT <= MAX_SUPPORTED_STACK_ALIGNMENT) |
1193 | { |
1194 | poly_int64 hwasan_orig_offset; |
1195 | if (hwasan_sanitize_stack_p ()) |
1196 | { |
1197 | /* There must be no tag granule "shared" between different |
1198 | objects. This means that no HWASAN_TAG_GRANULE_SIZE byte |
1199 | chunk can have more than one object in it. |
1200 | |
1201 | We ensure this by forcing the end of the last bit of data to |
1202 | be aligned to HWASAN_TAG_GRANULE_SIZE bytes here, and setting |
1203 | the start of each variable to be aligned to |
1204 | HWASAN_TAG_GRANULE_SIZE bytes in `align_local_variable`. |
1205 | |
1206 | We can't align just one of the start or end, since there are |
1207 | untagged things stored on the stack which we do not align to |
1208 | HWASAN_TAG_GRANULE_SIZE bytes. If we only aligned the start |
1209 | or the end of tagged objects then untagged objects could end |
1210 | up sharing the first granule of a tagged object or sharing the |
1211 | last granule of a tagged object respectively. */ |
1212 | hwasan_orig_offset = align_frame_offset (HWASAN_TAG_GRANULE_SIZE); |
1213 | gcc_assert (stack_vars[i].alignb >= HWASAN_TAG_GRANULE_SIZE); |
1214 | } |
1215 | /* ASAN description strings don't yet have a syntax for expressing |
1216 | polynomial offsets. */ |
1217 | HOST_WIDE_INT prev_offset; |
1218 | if (asan_sanitize_stack_p () |
1219 | && pred |
1220 | && frame_offset.is_constant (const_value: &prev_offset) |
1221 | && stack_vars[i].size.is_constant ()) |
1222 | { |
1223 | if (data->asan_vec.is_empty ()) |
1224 | { |
1225 | align_frame_offset (ASAN_RED_ZONE_SIZE); |
1226 | prev_offset = frame_offset.to_constant (); |
1227 | } |
1228 | prev_offset = align_base (base: prev_offset, |
1229 | ASAN_MIN_RED_ZONE_SIZE, |
1230 | align_up: !FRAME_GROWS_DOWNWARD); |
1231 | tree repr_decl = NULL_TREE; |
1232 | unsigned HOST_WIDE_INT size |
1233 | = asan_var_and_redzone_size (size: stack_vars[i].size.to_constant ()); |
1234 | if (data->asan_vec.is_empty ()) |
1235 | size = MAX (size, ASAN_RED_ZONE_SIZE); |
1236 | |
1237 | unsigned HOST_WIDE_INT alignment = MAX (alignb, |
1238 | ASAN_MIN_RED_ZONE_SIZE); |
1239 | offset = alloc_stack_frame_space (size, align: alignment); |
1240 | |
1241 | data->asan_vec.safe_push (obj: prev_offset); |
1242 | /* Allocating a constant amount of space from a constant |
1243 | starting offset must give a constant result. */ |
1244 | data->asan_vec.safe_push (obj: (offset + stack_vars[i].size) |
1245 | .to_constant ()); |
1246 | /* Find best representative of the partition. |
1247 | Prefer those with DECL_NAME, even better |
1248 | satisfying asan_protect_stack_decl predicate. */ |
1249 | for (j = i; j != EOC; j = stack_vars[j].next) |
1250 | if (asan_protect_stack_decl (decl: stack_vars[j].decl) |
1251 | && DECL_NAME (stack_vars[j].decl)) |
1252 | { |
1253 | repr_decl = stack_vars[j].decl; |
1254 | break; |
1255 | } |
1256 | else if (repr_decl == NULL_TREE |
1257 | && DECL_P (stack_vars[j].decl) |
1258 | && DECL_NAME (stack_vars[j].decl)) |
1259 | repr_decl = stack_vars[j].decl; |
1260 | if (repr_decl == NULL_TREE) |
1261 | repr_decl = stack_vars[i].decl; |
1262 | data->asan_decl_vec.safe_push (obj: repr_decl); |
1263 | |
1264 | /* Make sure a representative is unpoison if another |
1265 | variable in the partition is handled by |
1266 | use-after-scope sanitization. */ |
1267 | if (asan_handled_variables != NULL |
1268 | && !asan_handled_variables->contains (k: repr_decl)) |
1269 | { |
1270 | for (j = i; j != EOC; j = stack_vars[j].next) |
1271 | if (asan_handled_variables->contains (k: stack_vars[j].decl)) |
1272 | break; |
1273 | if (j != EOC) |
1274 | asan_handled_variables->add (k: repr_decl); |
1275 | } |
1276 | |
1277 | data->asan_alignb = MAX (data->asan_alignb, alignb); |
1278 | if (data->asan_base == NULL) |
1279 | data->asan_base = gen_reg_rtx (Pmode); |
1280 | base = data->asan_base; |
1281 | |
1282 | if (!STRICT_ALIGNMENT) |
1283 | base_align = crtl->max_used_stack_slot_alignment; |
1284 | else |
1285 | base_align = MAX (crtl->max_used_stack_slot_alignment, |
1286 | GET_MODE_ALIGNMENT (SImode) |
1287 | << ASAN_SHADOW_SHIFT); |
1288 | } |
1289 | else |
1290 | { |
1291 | offset = alloc_stack_frame_space (size: stack_vars[i].size, align: alignb); |
1292 | base_align = crtl->max_used_stack_slot_alignment; |
1293 | |
1294 | if (hwasan_sanitize_stack_p ()) |
1295 | { |
1296 | /* Align again since the point of this alignment is to handle |
1297 | the "end" of the object (i.e. smallest address after the |
1298 | stack object). For FRAME_GROWS_DOWNWARD that requires |
1299 | aligning the stack before allocating, but for a frame that |
1300 | grows upwards that requires aligning the stack after |
1301 | allocation. |
1302 | |
1303 | Use `frame_offset` to record the offset value rather than |
1304 | `offset` since the `frame_offset` describes the extent |
1305 | allocated for this particular variable while `offset` |
1306 | describes the address that this variable starts at. */ |
1307 | align_frame_offset (HWASAN_TAG_GRANULE_SIZE); |
1308 | hwasan_record_stack_var (virtual_stack_vars_rtx, base, |
1309 | hwasan_orig_offset, frame_offset); |
1310 | } |
1311 | } |
1312 | } |
1313 | else |
1314 | { |
1315 | /* Large alignment is only processed in the last pass. */ |
1316 | if (pred) |
1317 | continue; |
1318 | |
1319 | /* If there were any variables requiring "large" alignment, allocate |
1320 | space. */ |
1321 | if (maybe_ne (a: large_size, b: 0U) && ! large_allocation_done) |
1322 | { |
1323 | poly_int64 loffset; |
1324 | rtx large_allocsize; |
1325 | |
1326 | large_allocsize = gen_int_mode (large_size, Pmode); |
1327 | get_dynamic_stack_size (&large_allocsize, 0, large_align, NULL); |
1328 | loffset = alloc_stack_frame_space |
1329 | (size: rtx_to_poly_int64 (x: large_allocsize), |
1330 | PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT); |
1331 | large_base = get_dynamic_stack_base (loffset, large_align, base); |
1332 | large_allocation_done = true; |
1333 | } |
1334 | |
1335 | gcc_assert (large_base != NULL); |
1336 | large_alloc = aligned_upper_bound (value: large_alloc, align: alignb); |
1337 | offset = large_alloc; |
1338 | large_alloc += stack_vars[i].size; |
1339 | if (hwasan_sanitize_stack_p ()) |
1340 | { |
1341 | /* An object with a large alignment requirement means that the |
1342 | alignment requirement is greater than the required alignment |
1343 | for tags. */ |
1344 | if (!large_untagged_base) |
1345 | large_untagged_base |
1346 | = targetm.memtag.untagged_pointer (large_base, NULL_RTX); |
1347 | /* Ensure the end of the variable is also aligned correctly. */ |
1348 | poly_int64 align_again |
1349 | = aligned_upper_bound (value: large_alloc, HWASAN_TAG_GRANULE_SIZE); |
1350 | /* For large allocations we always allocate a chunk of space |
1351 | (which is addressed by large_untagged_base/large_base) and |
1352 | then use positive offsets from that. Hence the farthest |
1353 | offset is `align_again` and the nearest offset from the base |
1354 | is `offset`. */ |
1355 | hwasan_record_stack_var (large_untagged_base, large_base, |
1356 | offset, align_again); |
1357 | } |
1358 | |
1359 | base = large_base; |
1360 | base_align = large_align; |
1361 | } |
1362 | |
1363 | /* Create rtl for each variable based on their location within the |
1364 | partition. */ |
1365 | for (j = i; j != EOC; j = stack_vars[j].next) |
1366 | { |
1367 | expand_one_stack_var_at (decl: stack_vars[j].decl, |
1368 | base, base_align, offset); |
1369 | } |
1370 | if (hwasan_sanitize_stack_p ()) |
1371 | hwasan_increment_frame_tag (); |
1372 | } |
1373 | |
1374 | gcc_assert (known_eq (large_alloc, large_size)); |
1375 | } |
1376 | |
1377 | /* Take into account all sizes of partitions and reset DECL_RTLs. */ |
1378 | static poly_uint64 |
1379 | account_stack_vars (void) |
1380 | { |
1381 | size_t si, j, i, n = stack_vars_num; |
1382 | poly_uint64 size = 0; |
1383 | |
1384 | for (si = 0; si < n; ++si) |
1385 | { |
1386 | i = stack_vars_sorted[si]; |
1387 | |
1388 | /* Skip variables that aren't partition representatives, for now. */ |
1389 | if (stack_vars[i].representative != i) |
1390 | continue; |
1391 | |
1392 | size += stack_vars[i].size; |
1393 | for (j = i; j != EOC; j = stack_vars[j].next) |
1394 | set_rtl (t: stack_vars[j].decl, NULL); |
1395 | } |
1396 | return size; |
1397 | } |
1398 | |
1399 | /* Record the RTL assignment X for the default def of PARM. */ |
1400 | |
1401 | extern void |
1402 | set_parm_rtl (tree parm, rtx x) |
1403 | { |
1404 | gcc_assert (TREE_CODE (parm) == PARM_DECL |
1405 | || TREE_CODE (parm) == RESULT_DECL); |
1406 | |
1407 | if (x && !MEM_P (x)) |
1408 | { |
1409 | unsigned int align = MINIMUM_ALIGNMENT (TREE_TYPE (parm), |
1410 | TYPE_MODE (TREE_TYPE (parm)), |
1411 | TYPE_ALIGN (TREE_TYPE (parm))); |
1412 | |
1413 | /* If the variable alignment is very large we'll dynamicaly |
1414 | allocate it, which means that in-frame portion is just a |
1415 | pointer. ??? We've got a pseudo for sure here, do we |
1416 | actually dynamically allocate its spilling area if needed? |
1417 | ??? Isn't it a problem when Pmode alignment also exceeds |
1418 | MAX_SUPPORTED_STACK_ALIGNMENT, as can happen on cris and lm32? */ |
1419 | if (align > MAX_SUPPORTED_STACK_ALIGNMENT) |
1420 | align = GET_MODE_ALIGNMENT (Pmode); |
1421 | |
1422 | record_alignment_for_reg_var (align); |
1423 | } |
1424 | |
1425 | tree ssa = ssa_default_def (cfun, parm); |
1426 | if (!ssa) |
1427 | return set_rtl (t: parm, x); |
1428 | |
1429 | int part = var_to_partition (map: SA.map, var: ssa); |
1430 | gcc_assert (part != NO_PARTITION); |
1431 | |
1432 | bool changed = bitmap_bit_p (SA.partitions_for_parm_default_defs, part); |
1433 | gcc_assert (changed); |
1434 | |
1435 | set_rtl (t: ssa, x); |
1436 | gcc_assert (DECL_RTL (parm) == x); |
1437 | } |
1438 | |
1439 | /* A subroutine of expand_one_var. Called to immediately assign rtl |
1440 | to a variable to be allocated in the stack frame. */ |
1441 | |
1442 | static void |
1443 | expand_one_stack_var_1 (tree var) |
1444 | { |
1445 | poly_uint64 size; |
1446 | poly_int64 offset; |
1447 | unsigned byte_align; |
1448 | |
1449 | if (TREE_CODE (var) == SSA_NAME) |
1450 | { |
1451 | tree type = TREE_TYPE (var); |
1452 | size = tree_to_poly_uint64 (TYPE_SIZE_UNIT (type)); |
1453 | } |
1454 | else |
1455 | size = tree_to_poly_uint64 (DECL_SIZE_UNIT (var)); |
1456 | |
1457 | byte_align = align_local_variable (decl: var, really_expand: true); |
1458 | |
1459 | /* We handle highly aligned variables in expand_stack_vars. */ |
1460 | gcc_assert (byte_align * BITS_PER_UNIT <= MAX_SUPPORTED_STACK_ALIGNMENT); |
1461 | |
1462 | rtx base; |
1463 | if (hwasan_sanitize_stack_p ()) |
1464 | { |
1465 | /* Allocate zero bytes to align the stack. */ |
1466 | poly_int64 hwasan_orig_offset |
1467 | = align_frame_offset (HWASAN_TAG_GRANULE_SIZE); |
1468 | offset = alloc_stack_frame_space (size, align: byte_align); |
1469 | align_frame_offset (HWASAN_TAG_GRANULE_SIZE); |
1470 | base = hwasan_frame_base (); |
1471 | /* Use `frame_offset` to automatically account for machines where the |
1472 | frame grows upwards. |
1473 | |
1474 | `offset` will always point to the "start" of the stack object, which |
1475 | will be the smallest address, for ! FRAME_GROWS_DOWNWARD this is *not* |
1476 | the "furthest" offset from the base delimiting the current stack |
1477 | object. `frame_offset` will always delimit the extent that the frame. |
1478 | */ |
1479 | hwasan_record_stack_var (virtual_stack_vars_rtx, base, |
1480 | hwasan_orig_offset, frame_offset); |
1481 | } |
1482 | else |
1483 | { |
1484 | offset = alloc_stack_frame_space (size, align: byte_align); |
1485 | base = virtual_stack_vars_rtx; |
1486 | } |
1487 | |
1488 | expand_one_stack_var_at (decl: var, base, |
1489 | crtl->max_used_stack_slot_alignment, offset); |
1490 | |
1491 | if (hwasan_sanitize_stack_p ()) |
1492 | hwasan_increment_frame_tag (); |
1493 | } |
1494 | |
1495 | /* Wrapper for expand_one_stack_var_1 that checks SSA_NAMEs are |
1496 | already assigned some MEM. */ |
1497 | |
1498 | static void |
1499 | expand_one_stack_var (tree var) |
1500 | { |
1501 | if (TREE_CODE (var) == SSA_NAME) |
1502 | { |
1503 | int part = var_to_partition (map: SA.map, var); |
1504 | if (part != NO_PARTITION) |
1505 | { |
1506 | rtx x = SA.partition_to_pseudo[part]; |
1507 | gcc_assert (x); |
1508 | gcc_assert (MEM_P (x)); |
1509 | return; |
1510 | } |
1511 | } |
1512 | |
1513 | return expand_one_stack_var_1 (var); |
1514 | } |
1515 | |
1516 | /* A subroutine of expand_one_var. Called to assign rtl to a VAR_DECL |
1517 | that will reside in a hard register. */ |
1518 | |
1519 | static void |
1520 | expand_one_hard_reg_var (tree var) |
1521 | { |
1522 | rest_of_decl_compilation (var, 0, 0); |
1523 | } |
1524 | |
1525 | /* Record the alignment requirements of some variable assigned to a |
1526 | pseudo. */ |
1527 | |
1528 | static void |
1529 | record_alignment_for_reg_var (unsigned int align) |
1530 | { |
1531 | if (SUPPORTS_STACK_ALIGNMENT |
1532 | && crtl->stack_alignment_estimated < align) |
1533 | { |
1534 | /* stack_alignment_estimated shouldn't change after stack |
1535 | realign decision made */ |
1536 | gcc_assert (!crtl->stack_realign_processed); |
1537 | crtl->stack_alignment_estimated = align; |
1538 | } |
1539 | |
1540 | /* stack_alignment_needed > PREFERRED_STACK_BOUNDARY is permitted. |
1541 | So here we only make sure stack_alignment_needed >= align. */ |
1542 | if (crtl->stack_alignment_needed < align) |
1543 | crtl->stack_alignment_needed = align; |
1544 | if (crtl->max_used_stack_slot_alignment < align) |
1545 | crtl->max_used_stack_slot_alignment = align; |
1546 | } |
1547 | |
1548 | /* Create RTL for an SSA partition. */ |
1549 | |
1550 | static void |
1551 | expand_one_ssa_partition (tree var) |
1552 | { |
1553 | int part = var_to_partition (map: SA.map, var); |
1554 | gcc_assert (part != NO_PARTITION); |
1555 | |
1556 | if (SA.partition_to_pseudo[part]) |
1557 | return; |
1558 | |
1559 | unsigned int align = MINIMUM_ALIGNMENT (TREE_TYPE (var), |
1560 | TYPE_MODE (TREE_TYPE (var)), |
1561 | TYPE_ALIGN (TREE_TYPE (var))); |
1562 | |
1563 | /* If the variable alignment is very large we'll dynamicaly allocate |
1564 | it, which means that in-frame portion is just a pointer. */ |
1565 | if (align > MAX_SUPPORTED_STACK_ALIGNMENT) |
1566 | align = GET_MODE_ALIGNMENT (Pmode); |
1567 | |
1568 | record_alignment_for_reg_var (align); |
1569 | |
1570 | if (!use_register_for_decl (var)) |
1571 | { |
1572 | if (defer_stack_allocation (var, true)) |
1573 | add_stack_var (decl: var, really_expand: true); |
1574 | else |
1575 | expand_one_stack_var_1 (var); |
1576 | return; |
1577 | } |
1578 | |
1579 | machine_mode reg_mode = promote_ssa_mode (var, NULL); |
1580 | rtx x = gen_reg_rtx (reg_mode); |
1581 | |
1582 | set_rtl (t: var, x); |
1583 | |
1584 | /* For a promoted variable, X will not be used directly but wrapped in a |
1585 | SUBREG with SUBREG_PROMOTED_VAR_P set, which means that the RTL land |
1586 | will assume that its upper bits can be inferred from its lower bits. |
1587 | Therefore, if X isn't initialized on every path from the entry, then |
1588 | we must do it manually in order to fulfill the above assumption. */ |
1589 | if (reg_mode != TYPE_MODE (TREE_TYPE (var)) |
1590 | && bitmap_bit_p (SA.partitions_for_undefined_values, part)) |
1591 | emit_move_insn (x, CONST0_RTX (reg_mode)); |
1592 | } |
1593 | |
1594 | /* Record the association between the RTL generated for partition PART |
1595 | and the underlying variable of the SSA_NAME VAR. */ |
1596 | |
1597 | static void |
1598 | adjust_one_expanded_partition_var (tree var) |
1599 | { |
1600 | if (!var) |
1601 | return; |
1602 | |
1603 | tree decl = SSA_NAME_VAR (var); |
1604 | |
1605 | int part = var_to_partition (map: SA.map, var); |
1606 | if (part == NO_PARTITION) |
1607 | return; |
1608 | |
1609 | rtx x = SA.partition_to_pseudo[part]; |
1610 | |
1611 | gcc_assert (x); |
1612 | |
1613 | set_rtl (t: var, x); |
1614 | |
1615 | if (!REG_P (x)) |
1616 | return; |
1617 | |
1618 | /* Note if the object is a user variable. */ |
1619 | if (decl && !DECL_ARTIFICIAL (decl)) |
1620 | mark_user_reg (x); |
1621 | |
1622 | if (POINTER_TYPE_P (decl ? TREE_TYPE (decl) : TREE_TYPE (var))) |
1623 | mark_reg_pointer (x, get_pointer_alignment (var)); |
1624 | } |
1625 | |
1626 | /* A subroutine of expand_one_var. Called to assign rtl to a VAR_DECL |
1627 | that will reside in a pseudo register. */ |
1628 | |
1629 | static void |
1630 | expand_one_register_var (tree var) |
1631 | { |
1632 | if (TREE_CODE (var) == SSA_NAME) |
1633 | { |
1634 | int part = var_to_partition (map: SA.map, var); |
1635 | if (part != NO_PARTITION) |
1636 | { |
1637 | rtx x = SA.partition_to_pseudo[part]; |
1638 | gcc_assert (x); |
1639 | gcc_assert (REG_P (x)); |
1640 | return; |
1641 | } |
1642 | gcc_unreachable (); |
1643 | } |
1644 | |
1645 | tree decl = var; |
1646 | tree type = TREE_TYPE (decl); |
1647 | machine_mode reg_mode = promote_decl_mode (decl, NULL); |
1648 | rtx x = gen_reg_rtx (reg_mode); |
1649 | |
1650 | set_rtl (t: var, x); |
1651 | |
1652 | /* Note if the object is a user variable. */ |
1653 | if (!DECL_ARTIFICIAL (decl)) |
1654 | mark_user_reg (x); |
1655 | |
1656 | if (POINTER_TYPE_P (type)) |
1657 | mark_reg_pointer (x, get_pointer_alignment (var)); |
1658 | } |
1659 | |
1660 | /* A subroutine of expand_one_var. Called to assign rtl to a VAR_DECL that |
1661 | has some associated error, e.g. its type is error-mark. We just need |
1662 | to pick something that won't crash the rest of the compiler. */ |
1663 | |
1664 | static void |
1665 | expand_one_error_var (tree var) |
1666 | { |
1667 | machine_mode mode = DECL_MODE (var); |
1668 | rtx x; |
1669 | |
1670 | if (mode == BLKmode) |
1671 | x = gen_rtx_MEM (BLKmode, const0_rtx); |
1672 | else if (mode == VOIDmode) |
1673 | x = const0_rtx; |
1674 | else |
1675 | x = gen_reg_rtx (mode); |
1676 | |
1677 | SET_DECL_RTL (var, x); |
1678 | } |
1679 | |
1680 | /* A subroutine of expand_one_var. VAR is a variable that will be |
1681 | allocated to the local stack frame. Return true if we wish to |
1682 | add VAR to STACK_VARS so that it will be coalesced with other |
1683 | variables. Return false to allocate VAR immediately. |
1684 | |
1685 | This function is used to reduce the number of variables considered |
1686 | for coalescing, which reduces the size of the quadratic problem. */ |
1687 | |
1688 | static bool |
1689 | defer_stack_allocation (tree var, bool toplevel) |
1690 | { |
1691 | tree size_unit = TREE_CODE (var) == SSA_NAME |
1692 | ? TYPE_SIZE_UNIT (TREE_TYPE (var)) |
1693 | : DECL_SIZE_UNIT (var); |
1694 | poly_uint64 size; |
1695 | |
1696 | /* Whether the variable is small enough for immediate allocation not to be |
1697 | a problem with regard to the frame size. */ |
1698 | bool smallish |
1699 | = (poly_int_tree_p (t: size_unit, value: &size) |
1700 | && (estimated_poly_value (x: size) |
1701 | < param_min_size_for_stack_sharing)); |
1702 | |
1703 | /* If stack protection is enabled, *all* stack variables must be deferred, |
1704 | so that we can re-order the strings to the top of the frame. |
1705 | Similarly for Address Sanitizer. */ |
1706 | if (flag_stack_protect || asan_sanitize_stack_p ()) |
1707 | return true; |
1708 | |
1709 | unsigned int align = TREE_CODE (var) == SSA_NAME |
1710 | ? TYPE_ALIGN (TREE_TYPE (var)) |
1711 | : DECL_ALIGN (var); |
1712 | |
1713 | /* We handle "large" alignment via dynamic allocation. We want to handle |
1714 | this extra complication in only one place, so defer them. */ |
1715 | if (align > MAX_SUPPORTED_STACK_ALIGNMENT) |
1716 | return true; |
1717 | |
1718 | bool ignored = TREE_CODE (var) == SSA_NAME |
1719 | ? !SSAVAR (var) || DECL_IGNORED_P (SSA_NAME_VAR (var)) |
1720 | : DECL_IGNORED_P (var); |
1721 | |
1722 | /* When optimization is enabled, DECL_IGNORED_P variables originally scoped |
1723 | might be detached from their block and appear at toplevel when we reach |
1724 | here. We want to coalesce them with variables from other blocks when |
1725 | the immediate contribution to the frame size would be noticeable. */ |
1726 | if (toplevel && optimize > 0 && ignored && !smallish) |
1727 | return true; |
1728 | |
1729 | /* Variables declared in the outermost scope automatically conflict |
1730 | with every other variable. The only reason to want to defer them |
1731 | at all is that, after sorting, we can more efficiently pack |
1732 | small variables in the stack frame. Continue to defer at -O2. */ |
1733 | if (toplevel && optimize < 2) |
1734 | return false; |
1735 | |
1736 | /* Without optimization, *most* variables are allocated from the |
1737 | stack, which makes the quadratic problem large exactly when we |
1738 | want compilation to proceed as quickly as possible. On the |
1739 | other hand, we don't want the function's stack frame size to |
1740 | get completely out of hand. So we avoid adding scalars and |
1741 | "small" aggregates to the list at all. */ |
1742 | if (optimize == 0 && smallish) |
1743 | return false; |
1744 | |
1745 | return true; |
1746 | } |
1747 | |
1748 | /* A subroutine of expand_used_vars. Expand one variable according to |
1749 | its flavor. Variables to be placed on the stack are not actually |
1750 | expanded yet, merely recorded. |
1751 | When REALLY_EXPAND is false, only add stack values to be allocated. |
1752 | Return stack usage this variable is supposed to take. |
1753 | */ |
1754 | |
1755 | static poly_uint64 |
1756 | expand_one_var (tree var, bool toplevel, bool really_expand, |
1757 | bitmap forced_stack_var = NULL) |
1758 | { |
1759 | unsigned int align = BITS_PER_UNIT; |
1760 | tree origvar = var; |
1761 | |
1762 | var = SSAVAR (var); |
1763 | |
1764 | if (TREE_TYPE (var) != error_mark_node && VAR_P (var)) |
1765 | { |
1766 | if (is_global_var (t: var)) |
1767 | return 0; |
1768 | |
1769 | /* Because we don't know if VAR will be in register or on stack, |
1770 | we conservatively assume it will be on stack even if VAR is |
1771 | eventually put into register after RA pass. For non-automatic |
1772 | variables, which won't be on stack, we collect alignment of |
1773 | type and ignore user specified alignment. Similarly for |
1774 | SSA_NAMEs for which use_register_for_decl returns true. */ |
1775 | if (TREE_STATIC (var) |
1776 | || DECL_EXTERNAL (var) |
1777 | || (TREE_CODE (origvar) == SSA_NAME && use_register_for_decl (var))) |
1778 | align = MINIMUM_ALIGNMENT (TREE_TYPE (var), |
1779 | TYPE_MODE (TREE_TYPE (var)), |
1780 | TYPE_ALIGN (TREE_TYPE (var))); |
1781 | else if (DECL_HAS_VALUE_EXPR_P (var) |
1782 | || (DECL_RTL_SET_P (var) && MEM_P (DECL_RTL (var)))) |
1783 | /* Don't consider debug only variables with DECL_HAS_VALUE_EXPR_P set |
1784 | or variables which were assigned a stack slot already by |
1785 | expand_one_stack_var_at - in the latter case DECL_ALIGN has been |
1786 | changed from the offset chosen to it. */ |
1787 | align = crtl->stack_alignment_estimated; |
1788 | else |
1789 | align = MINIMUM_ALIGNMENT (var, DECL_MODE (var), DECL_ALIGN (var)); |
1790 | |
1791 | /* If the variable alignment is very large we'll dynamicaly allocate |
1792 | it, which means that in-frame portion is just a pointer. */ |
1793 | if (align > MAX_SUPPORTED_STACK_ALIGNMENT) |
1794 | align = GET_MODE_ALIGNMENT (Pmode); |
1795 | } |
1796 | |
1797 | record_alignment_for_reg_var (align); |
1798 | |
1799 | poly_uint64 size; |
1800 | if (TREE_CODE (origvar) == SSA_NAME) |
1801 | { |
1802 | gcc_assert (!VAR_P (var) |
1803 | || (!DECL_EXTERNAL (var) |
1804 | && !DECL_HAS_VALUE_EXPR_P (var) |
1805 | && !TREE_STATIC (var) |
1806 | && TREE_TYPE (var) != error_mark_node |
1807 | && !DECL_HARD_REGISTER (var) |
1808 | && really_expand)); |
1809 | } |
1810 | if (!VAR_P (var) && TREE_CODE (origvar) != SSA_NAME) |
1811 | ; |
1812 | else if (DECL_EXTERNAL (var)) |
1813 | ; |
1814 | else if (DECL_HAS_VALUE_EXPR_P (var)) |
1815 | ; |
1816 | else if (TREE_STATIC (var)) |
1817 | ; |
1818 | else if (TREE_CODE (origvar) != SSA_NAME && DECL_RTL_SET_P (var)) |
1819 | ; |
1820 | else if (TREE_TYPE (var) == error_mark_node) |
1821 | { |
1822 | if (really_expand) |
1823 | expand_one_error_var (var); |
1824 | } |
1825 | else if (VAR_P (var) && DECL_HARD_REGISTER (var)) |
1826 | { |
1827 | if (really_expand) |
1828 | { |
1829 | expand_one_hard_reg_var (var); |
1830 | if (!DECL_HARD_REGISTER (var)) |
1831 | /* Invalid register specification. */ |
1832 | expand_one_error_var (var); |
1833 | } |
1834 | } |
1835 | else if (use_register_for_decl (var) |
1836 | && (!forced_stack_var |
1837 | || !bitmap_bit_p (forced_stack_var, DECL_UID (var)))) |
1838 | { |
1839 | if (really_expand) |
1840 | expand_one_register_var (var: origvar); |
1841 | } |
1842 | else if (!poly_int_tree_p (DECL_SIZE_UNIT (var), value: &size) |
1843 | || !valid_constant_size_p (DECL_SIZE_UNIT (var))) |
1844 | { |
1845 | /* Reject variables which cover more than half of the address-space. */ |
1846 | if (really_expand) |
1847 | { |
1848 | if (DECL_NONLOCAL_FRAME (var)) |
1849 | error_at (DECL_SOURCE_LOCATION (current_function_decl), |
1850 | "total size of local objects is too large" ); |
1851 | else |
1852 | error_at (DECL_SOURCE_LOCATION (var), |
1853 | "size of variable %q+D is too large" , var); |
1854 | expand_one_error_var (var); |
1855 | } |
1856 | } |
1857 | else if (defer_stack_allocation (var, toplevel)) |
1858 | add_stack_var (decl: origvar, really_expand); |
1859 | else |
1860 | { |
1861 | if (really_expand) |
1862 | { |
1863 | if (lookup_attribute (attr_name: "naked" , |
1864 | DECL_ATTRIBUTES (current_function_decl))) |
1865 | error ("cannot allocate stack for variable %q+D, naked function" , |
1866 | var); |
1867 | |
1868 | expand_one_stack_var (var: origvar); |
1869 | } |
1870 | return size; |
1871 | } |
1872 | return 0; |
1873 | } |
1874 | |
1875 | /* A subroutine of expand_used_vars. Walk down through the BLOCK tree |
1876 | expanding variables. Those variables that can be put into registers |
1877 | are allocated pseudos; those that can't are put on the stack. |
1878 | |
1879 | TOPLEVEL is true if this is the outermost BLOCK. */ |
1880 | |
1881 | static void |
1882 | expand_used_vars_for_block (tree block, bool toplevel, bitmap forced_stack_vars) |
1883 | { |
1884 | tree t; |
1885 | |
1886 | /* Expand all variables at this level. */ |
1887 | for (t = BLOCK_VARS (block); t ; t = DECL_CHAIN (t)) |
1888 | if (TREE_USED (t) |
1889 | && ((!VAR_P (t) && TREE_CODE (t) != RESULT_DECL) |
1890 | || !DECL_NONSHAREABLE (t))) |
1891 | expand_one_var (var: t, toplevel, really_expand: true, forced_stack_var: forced_stack_vars); |
1892 | |
1893 | /* Expand all variables at containing levels. */ |
1894 | for (t = BLOCK_SUBBLOCKS (block); t ; t = BLOCK_CHAIN (t)) |
1895 | expand_used_vars_for_block (block: t, toplevel: false, forced_stack_vars); |
1896 | } |
1897 | |
1898 | /* A subroutine of expand_used_vars. Walk down through the BLOCK tree |
1899 | and clear TREE_USED on all local variables. */ |
1900 | |
1901 | static void |
1902 | clear_tree_used (tree block) |
1903 | { |
1904 | tree t; |
1905 | |
1906 | for (t = BLOCK_VARS (block); t ; t = DECL_CHAIN (t)) |
1907 | /* if (!TREE_STATIC (t) && !DECL_EXTERNAL (t)) */ |
1908 | if ((!VAR_P (t) && TREE_CODE (t) != RESULT_DECL) |
1909 | || !DECL_NONSHAREABLE (t)) |
1910 | TREE_USED (t) = 0; |
1911 | |
1912 | for (t = BLOCK_SUBBLOCKS (block); t ; t = BLOCK_CHAIN (t)) |
1913 | clear_tree_used (block: t); |
1914 | } |
1915 | |
1916 | /* Examine TYPE and determine a bit mask of the following features. */ |
1917 | |
1918 | #define SPCT_HAS_LARGE_CHAR_ARRAY 1 |
1919 | #define SPCT_HAS_SMALL_CHAR_ARRAY 2 |
1920 | #define SPCT_HAS_ARRAY 4 |
1921 | #define SPCT_HAS_AGGREGATE 8 |
1922 | |
1923 | static unsigned int |
1924 | stack_protect_classify_type (tree type) |
1925 | { |
1926 | unsigned int ret = 0; |
1927 | tree t; |
1928 | |
1929 | switch (TREE_CODE (type)) |
1930 | { |
1931 | case ARRAY_TYPE: |
1932 | t = TYPE_MAIN_VARIANT (TREE_TYPE (type)); |
1933 | if (t == char_type_node |
1934 | || t == signed_char_type_node |
1935 | || t == unsigned_char_type_node) |
1936 | { |
1937 | unsigned HOST_WIDE_INT max = param_ssp_buffer_size; |
1938 | unsigned HOST_WIDE_INT len; |
1939 | |
1940 | if (!TYPE_SIZE_UNIT (type) |
1941 | || !tree_fits_uhwi_p (TYPE_SIZE_UNIT (type))) |
1942 | len = max; |
1943 | else |
1944 | len = tree_to_uhwi (TYPE_SIZE_UNIT (type)); |
1945 | |
1946 | if (len < max) |
1947 | ret = SPCT_HAS_SMALL_CHAR_ARRAY | SPCT_HAS_ARRAY; |
1948 | else |
1949 | ret = SPCT_HAS_LARGE_CHAR_ARRAY | SPCT_HAS_ARRAY; |
1950 | } |
1951 | else |
1952 | ret = SPCT_HAS_ARRAY; |
1953 | break; |
1954 | |
1955 | case UNION_TYPE: |
1956 | case QUAL_UNION_TYPE: |
1957 | case RECORD_TYPE: |
1958 | ret = SPCT_HAS_AGGREGATE; |
1959 | for (t = TYPE_FIELDS (type); t ; t = TREE_CHAIN (t)) |
1960 | if (TREE_CODE (t) == FIELD_DECL) |
1961 | ret |= stack_protect_classify_type (TREE_TYPE (t)); |
1962 | break; |
1963 | |
1964 | default: |
1965 | break; |
1966 | } |
1967 | |
1968 | return ret; |
1969 | } |
1970 | |
1971 | /* Return nonzero if DECL should be segregated into the "vulnerable" upper |
1972 | part of the local stack frame. Remember if we ever return nonzero for |
1973 | any variable in this function. The return value is the phase number in |
1974 | which the variable should be allocated. */ |
1975 | |
1976 | static int |
1977 | stack_protect_decl_phase (tree decl) |
1978 | { |
1979 | unsigned int bits = stack_protect_classify_type (TREE_TYPE (decl)); |
1980 | int ret = 0; |
1981 | |
1982 | if (bits & SPCT_HAS_SMALL_CHAR_ARRAY) |
1983 | has_short_buffer = true; |
1984 | |
1985 | tree attribs = DECL_ATTRIBUTES (current_function_decl); |
1986 | if (!lookup_attribute (attr_name: "no_stack_protector" , list: attribs) |
1987 | && (flag_stack_protect == SPCT_FLAG_ALL |
1988 | || flag_stack_protect == SPCT_FLAG_STRONG |
1989 | || (flag_stack_protect == SPCT_FLAG_EXPLICIT |
1990 | && lookup_attribute (attr_name: "stack_protect" , list: attribs)))) |
1991 | { |
1992 | if ((bits & (SPCT_HAS_SMALL_CHAR_ARRAY | SPCT_HAS_LARGE_CHAR_ARRAY)) |
1993 | && !(bits & SPCT_HAS_AGGREGATE)) |
1994 | ret = 1; |
1995 | else if (bits & SPCT_HAS_ARRAY) |
1996 | ret = 2; |
1997 | } |
1998 | else |
1999 | ret = (bits & SPCT_HAS_LARGE_CHAR_ARRAY) != 0; |
2000 | |
2001 | if (ret) |
2002 | has_protected_decls = true; |
2003 | |
2004 | return ret; |
2005 | } |
2006 | |
2007 | /* Two helper routines that check for phase 1 and phase 2. These are used |
2008 | as callbacks for expand_stack_vars. */ |
2009 | |
2010 | static bool |
2011 | stack_protect_decl_phase_1 (size_t i) |
2012 | { |
2013 | return stack_protect_decl_phase (decl: stack_vars[i].decl) == 1; |
2014 | } |
2015 | |
2016 | static bool |
2017 | stack_protect_decl_phase_2 (size_t i) |
2018 | { |
2019 | return stack_protect_decl_phase (decl: stack_vars[i].decl) == 2; |
2020 | } |
2021 | |
2022 | /* And helper function that checks for asan phase (with stack protector |
2023 | it is phase 3). This is used as callback for expand_stack_vars. |
2024 | Returns true if any of the vars in the partition need to be protected. */ |
2025 | |
2026 | static bool |
2027 | asan_decl_phase_3 (size_t i) |
2028 | { |
2029 | while (i != EOC) |
2030 | { |
2031 | if (asan_protect_stack_decl (decl: stack_vars[i].decl)) |
2032 | return true; |
2033 | i = stack_vars[i].next; |
2034 | } |
2035 | return false; |
2036 | } |
2037 | |
2038 | /* Ensure that variables in different stack protection phases conflict |
2039 | so that they are not merged and share the same stack slot. |
2040 | Return true if there are any address taken variables. */ |
2041 | |
2042 | static bool |
2043 | add_stack_protection_conflicts (void) |
2044 | { |
2045 | size_t i, j, n = stack_vars_num; |
2046 | unsigned char *phase; |
2047 | bool ret = false; |
2048 | |
2049 | phase = XNEWVEC (unsigned char, n); |
2050 | for (i = 0; i < n; ++i) |
2051 | { |
2052 | phase[i] = stack_protect_decl_phase (decl: stack_vars[i].decl); |
2053 | if (TREE_ADDRESSABLE (stack_vars[i].decl)) |
2054 | ret = true; |
2055 | } |
2056 | |
2057 | for (i = 0; i < n; ++i) |
2058 | { |
2059 | unsigned char ph_i = phase[i]; |
2060 | for (j = i + 1; j < n; ++j) |
2061 | if (ph_i != phase[j]) |
2062 | add_stack_var_conflict (x: i, y: j); |
2063 | } |
2064 | |
2065 | XDELETEVEC (phase); |
2066 | return ret; |
2067 | } |
2068 | |
2069 | /* Create a decl for the guard at the top of the stack frame. */ |
2070 | |
2071 | static void |
2072 | create_stack_guard (void) |
2073 | { |
2074 | tree guard = build_decl (DECL_SOURCE_LOCATION (current_function_decl), |
2075 | VAR_DECL, NULL, ptr_type_node); |
2076 | TREE_THIS_VOLATILE (guard) = 1; |
2077 | TREE_USED (guard) = 1; |
2078 | expand_one_stack_var (var: guard); |
2079 | crtl->stack_protect_guard = guard; |
2080 | } |
2081 | |
2082 | /* Prepare for expanding variables. */ |
2083 | static void |
2084 | init_vars_expansion (void) |
2085 | { |
2086 | /* Conflict bitmaps, and a few related temporary bitmaps, go here. */ |
2087 | bitmap_obstack_initialize (&stack_var_bitmap_obstack); |
2088 | |
2089 | /* A map from decl to stack partition. */ |
2090 | decl_to_stack_part = new hash_map<tree, size_t>; |
2091 | |
2092 | /* Initialize local stack smashing state. */ |
2093 | has_protected_decls = false; |
2094 | has_short_buffer = false; |
2095 | if (hwasan_sanitize_stack_p ()) |
2096 | hwasan_record_frame_init (); |
2097 | } |
2098 | |
2099 | /* Free up stack variable graph data. */ |
2100 | static void |
2101 | fini_vars_expansion (void) |
2102 | { |
2103 | bitmap_obstack_release (&stack_var_bitmap_obstack); |
2104 | if (stack_vars) |
2105 | XDELETEVEC (stack_vars); |
2106 | if (stack_vars_sorted) |
2107 | XDELETEVEC (stack_vars_sorted); |
2108 | stack_vars = NULL; |
2109 | stack_vars_sorted = NULL; |
2110 | stack_vars_alloc = stack_vars_num = 0; |
2111 | delete decl_to_stack_part; |
2112 | decl_to_stack_part = NULL; |
2113 | } |
2114 | |
2115 | /* Make a fair guess for the size of the stack frame of the function |
2116 | in NODE. This doesn't have to be exact, the result is only used in |
2117 | the inline heuristics. So we don't want to run the full stack var |
2118 | packing algorithm (which is quadratic in the number of stack vars). |
2119 | Instead, we calculate the total size of all stack vars. This turns |
2120 | out to be a pretty fair estimate -- packing of stack vars doesn't |
2121 | happen very often. */ |
2122 | |
2123 | HOST_WIDE_INT |
2124 | estimated_stack_frame_size (struct cgraph_node *node) |
2125 | { |
2126 | poly_int64 size = 0; |
2127 | size_t i; |
2128 | tree var; |
2129 | struct function *fn = DECL_STRUCT_FUNCTION (node->decl); |
2130 | |
2131 | push_cfun (new_cfun: fn); |
2132 | |
2133 | init_vars_expansion (); |
2134 | |
2135 | FOR_EACH_LOCAL_DECL (fn, i, var) |
2136 | if (auto_var_in_fn_p (var, fn->decl)) |
2137 | size += expand_one_var (var, toplevel: true, really_expand: false); |
2138 | |
2139 | if (stack_vars_num > 0) |
2140 | { |
2141 | /* Fake sorting the stack vars for account_stack_vars (). */ |
2142 | stack_vars_sorted = XNEWVEC (size_t, stack_vars_num); |
2143 | for (i = 0; i < stack_vars_num; ++i) |
2144 | stack_vars_sorted[i] = i; |
2145 | size += account_stack_vars (); |
2146 | } |
2147 | |
2148 | fini_vars_expansion (); |
2149 | pop_cfun (); |
2150 | return estimated_poly_value (x: size); |
2151 | } |
2152 | |
2153 | /* Check if the current function has calls that use a return slot. */ |
2154 | |
2155 | static bool |
2156 | stack_protect_return_slot_p () |
2157 | { |
2158 | basic_block bb; |
2159 | |
2160 | FOR_ALL_BB_FN (bb, cfun) |
2161 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); |
2162 | !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
2163 | { |
2164 | gimple *stmt = gsi_stmt (i: gsi); |
2165 | /* This assumes that calls to internal-only functions never |
2166 | use a return slot. */ |
2167 | if (is_gimple_call (gs: stmt) |
2168 | && !gimple_call_internal_p (gs: stmt) |
2169 | && aggregate_value_p (TREE_TYPE (gimple_call_fntype (stmt)), |
2170 | gimple_call_fndecl (gs: stmt))) |
2171 | return true; |
2172 | } |
2173 | return false; |
2174 | } |
2175 | |
2176 | /* Expand all variables used in the function. */ |
2177 | |
2178 | static rtx_insn * |
2179 | expand_used_vars (bitmap forced_stack_vars) |
2180 | { |
2181 | tree var, outer_block = DECL_INITIAL (current_function_decl); |
2182 | auto_vec<tree> maybe_local_decls; |
2183 | rtx_insn *var_end_seq = NULL; |
2184 | unsigned i; |
2185 | unsigned len; |
2186 | bool gen_stack_protect_signal = false; |
2187 | |
2188 | /* Compute the phase of the stack frame for this function. */ |
2189 | { |
2190 | int align = PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT; |
2191 | int off = targetm.starting_frame_offset () % align; |
2192 | frame_phase = off ? align - off : 0; |
2193 | } |
2194 | |
2195 | /* Set TREE_USED on all variables in the local_decls. */ |
2196 | FOR_EACH_LOCAL_DECL (cfun, i, var) |
2197 | TREE_USED (var) = 1; |
2198 | /* Clear TREE_USED on all variables associated with a block scope. */ |
2199 | clear_tree_used (DECL_INITIAL (current_function_decl)); |
2200 | |
2201 | init_vars_expansion (); |
2202 | |
2203 | if (targetm.use_pseudo_pic_reg ()) |
2204 | pic_offset_table_rtx = gen_reg_rtx (Pmode); |
2205 | |
2206 | for (i = 0; i < SA.map->num_partitions; i++) |
2207 | { |
2208 | if (bitmap_bit_p (SA.partitions_for_parm_default_defs, i)) |
2209 | continue; |
2210 | |
2211 | tree var = partition_to_var (map: SA.map, i); |
2212 | |
2213 | gcc_assert (!virtual_operand_p (var)); |
2214 | |
2215 | expand_one_ssa_partition (var); |
2216 | } |
2217 | |
2218 | if (flag_stack_protect == SPCT_FLAG_STRONG) |
2219 | gen_stack_protect_signal = stack_protect_return_slot_p (); |
2220 | |
2221 | /* At this point all variables on the local_decls with TREE_USED |
2222 | set are not associated with any block scope. Lay them out. */ |
2223 | |
2224 | len = vec_safe_length (cfun->local_decls); |
2225 | FOR_EACH_LOCAL_DECL (cfun, i, var) |
2226 | { |
2227 | bool expand_now = false; |
2228 | |
2229 | /* Expanded above already. */ |
2230 | if (is_gimple_reg (var)) |
2231 | { |
2232 | TREE_USED (var) = 0; |
2233 | goto next; |
2234 | } |
2235 | /* We didn't set a block for static or extern because it's hard |
2236 | to tell the difference between a global variable (re)declared |
2237 | in a local scope, and one that's really declared there to |
2238 | begin with. And it doesn't really matter much, since we're |
2239 | not giving them stack space. Expand them now. */ |
2240 | else if (TREE_STATIC (var) || DECL_EXTERNAL (var)) |
2241 | expand_now = true; |
2242 | |
2243 | /* Expand variables not associated with any block now. Those created by |
2244 | the optimizers could be live anywhere in the function. Those that |
2245 | could possibly have been scoped originally and detached from their |
2246 | block will have their allocation deferred so we coalesce them with |
2247 | others when optimization is enabled. */ |
2248 | else if (TREE_USED (var)) |
2249 | expand_now = true; |
2250 | |
2251 | /* Finally, mark all variables on the list as used. We'll use |
2252 | this in a moment when we expand those associated with scopes. */ |
2253 | TREE_USED (var) = 1; |
2254 | |
2255 | if (expand_now) |
2256 | expand_one_var (var, toplevel: true, really_expand: true, forced_stack_var: forced_stack_vars); |
2257 | |
2258 | next: |
2259 | if (DECL_ARTIFICIAL (var) && !DECL_IGNORED_P (var)) |
2260 | { |
2261 | rtx rtl = DECL_RTL_IF_SET (var); |
2262 | |
2263 | /* Keep artificial non-ignored vars in cfun->local_decls |
2264 | chain until instantiate_decls. */ |
2265 | if (rtl && (MEM_P (rtl) || GET_CODE (rtl) == CONCAT)) |
2266 | add_local_decl (cfun, d: var); |
2267 | else if (rtl == NULL_RTX) |
2268 | /* If rtl isn't set yet, which can happen e.g. with |
2269 | -fstack-protector, retry before returning from this |
2270 | function. */ |
2271 | maybe_local_decls.safe_push (obj: var); |
2272 | } |
2273 | } |
2274 | |
2275 | /* We duplicated some of the decls in CFUN->LOCAL_DECLS. |
2276 | |
2277 | +-----------------+-----------------+ |
2278 | | ...processed... | ...duplicates...| |
2279 | +-----------------+-----------------+ |
2280 | ^ |
2281 | +-- LEN points here. |
2282 | |
2283 | We just want the duplicates, as those are the artificial |
2284 | non-ignored vars that we want to keep until instantiate_decls. |
2285 | Move them down and truncate the array. */ |
2286 | if (!vec_safe_is_empty (cfun->local_decls)) |
2287 | cfun->local_decls->block_remove (ix: 0, len); |
2288 | |
2289 | /* At this point, all variables within the block tree with TREE_USED |
2290 | set are actually used by the optimized function. Lay them out. */ |
2291 | expand_used_vars_for_block (block: outer_block, toplevel: true, forced_stack_vars); |
2292 | |
2293 | tree attribs = DECL_ATTRIBUTES (current_function_decl); |
2294 | if (stack_vars_num > 0) |
2295 | { |
2296 | bool has_addressable_vars = false; |
2297 | |
2298 | add_scope_conflicts (); |
2299 | |
2300 | /* If stack protection is enabled, we don't share space between |
2301 | vulnerable data and non-vulnerable data. */ |
2302 | if (flag_stack_protect != 0 |
2303 | && !lookup_attribute (attr_name: "no_stack_protector" , list: attribs) |
2304 | && (flag_stack_protect != SPCT_FLAG_EXPLICIT |
2305 | || (flag_stack_protect == SPCT_FLAG_EXPLICIT |
2306 | && lookup_attribute (attr_name: "stack_protect" , list: attribs)))) |
2307 | has_addressable_vars = add_stack_protection_conflicts (); |
2308 | |
2309 | if (flag_stack_protect == SPCT_FLAG_STRONG && has_addressable_vars) |
2310 | gen_stack_protect_signal = true; |
2311 | |
2312 | /* Now that we have collected all stack variables, and have computed a |
2313 | minimal interference graph, attempt to save some stack space. */ |
2314 | partition_stack_vars (); |
2315 | if (dump_file) |
2316 | dump_stack_var_partition (); |
2317 | } |
2318 | |
2319 | |
2320 | if (!lookup_attribute (attr_name: "no_stack_protector" , list: attribs)) |
2321 | switch (flag_stack_protect) |
2322 | { |
2323 | case SPCT_FLAG_ALL: |
2324 | create_stack_guard (); |
2325 | break; |
2326 | |
2327 | case SPCT_FLAG_STRONG: |
2328 | if (gen_stack_protect_signal |
2329 | || cfun->calls_alloca |
2330 | || has_protected_decls |
2331 | || lookup_attribute (attr_name: "stack_protect" , list: attribs)) |
2332 | create_stack_guard (); |
2333 | break; |
2334 | |
2335 | case SPCT_FLAG_DEFAULT: |
2336 | if (cfun->calls_alloca |
2337 | || has_protected_decls |
2338 | || lookup_attribute (attr_name: "stack_protect" , list: attribs)) |
2339 | create_stack_guard (); |
2340 | break; |
2341 | |
2342 | case SPCT_FLAG_EXPLICIT: |
2343 | if (lookup_attribute (attr_name: "stack_protect" , list: attribs)) |
2344 | create_stack_guard (); |
2345 | break; |
2346 | |
2347 | default: |
2348 | break; |
2349 | } |
2350 | |
2351 | /* Assign rtl to each variable based on these partitions. */ |
2352 | if (stack_vars_num > 0) |
2353 | { |
2354 | class stack_vars_data data; |
2355 | |
2356 | data.asan_base = NULL_RTX; |
2357 | data.asan_alignb = 0; |
2358 | |
2359 | /* Reorder decls to be protected by iterating over the variables |
2360 | array multiple times, and allocating out of each phase in turn. */ |
2361 | /* ??? We could probably integrate this into the qsort we did |
2362 | earlier, such that we naturally see these variables first, |
2363 | and thus naturally allocate things in the right order. */ |
2364 | if (has_protected_decls) |
2365 | { |
2366 | /* Phase 1 contains only character arrays. */ |
2367 | expand_stack_vars (pred: stack_protect_decl_phase_1, data: &data); |
2368 | |
2369 | /* Phase 2 contains other kinds of arrays. */ |
2370 | if (!lookup_attribute (attr_name: "no_stack_protector" , list: attribs) |
2371 | && (flag_stack_protect == SPCT_FLAG_ALL |
2372 | || flag_stack_protect == SPCT_FLAG_STRONG |
2373 | || (flag_stack_protect == SPCT_FLAG_EXPLICIT |
2374 | && lookup_attribute (attr_name: "stack_protect" , list: attribs)))) |
2375 | expand_stack_vars (pred: stack_protect_decl_phase_2, data: &data); |
2376 | } |
2377 | |
2378 | if (asan_sanitize_stack_p ()) |
2379 | /* Phase 3, any partitions that need asan protection |
2380 | in addition to phase 1 and 2. */ |
2381 | expand_stack_vars (pred: asan_decl_phase_3, data: &data); |
2382 | |
2383 | /* ASAN description strings don't yet have a syntax for expressing |
2384 | polynomial offsets. */ |
2385 | HOST_WIDE_INT prev_offset; |
2386 | if (!data.asan_vec.is_empty () |
2387 | && frame_offset.is_constant (const_value: &prev_offset)) |
2388 | { |
2389 | HOST_WIDE_INT offset, sz, redzonesz; |
2390 | redzonesz = ASAN_RED_ZONE_SIZE; |
2391 | sz = data.asan_vec[0] - prev_offset; |
2392 | if (data.asan_alignb > ASAN_RED_ZONE_SIZE |
2393 | && data.asan_alignb <= 4096 |
2394 | && sz + ASAN_RED_ZONE_SIZE >= (int) data.asan_alignb) |
2395 | redzonesz = ((sz + ASAN_RED_ZONE_SIZE + data.asan_alignb - 1) |
2396 | & ~(data.asan_alignb - HOST_WIDE_INT_1)) - sz; |
2397 | /* Allocating a constant amount of space from a constant |
2398 | starting offset must give a constant result. */ |
2399 | offset = (alloc_stack_frame_space (size: redzonesz, ASAN_RED_ZONE_SIZE) |
2400 | .to_constant ()); |
2401 | data.asan_vec.safe_push (obj: prev_offset); |
2402 | data.asan_vec.safe_push (obj: offset); |
2403 | /* Leave space for alignment if STRICT_ALIGNMENT. */ |
2404 | if (STRICT_ALIGNMENT) |
2405 | alloc_stack_frame_space (size: (GET_MODE_ALIGNMENT (SImode) |
2406 | << ASAN_SHADOW_SHIFT) |
2407 | / BITS_PER_UNIT, align: 1); |
2408 | |
2409 | var_end_seq |
2410 | = asan_emit_stack_protection (virtual_stack_vars_rtx, |
2411 | data.asan_base, |
2412 | data.asan_alignb, |
2413 | data.asan_vec.address (), |
2414 | data.asan_decl_vec.address (), |
2415 | data.asan_vec.length ()); |
2416 | } |
2417 | |
2418 | expand_stack_vars (NULL, data: &data); |
2419 | } |
2420 | |
2421 | if (hwasan_sanitize_stack_p ()) |
2422 | hwasan_emit_prologue (); |
2423 | if (asan_sanitize_allocas_p () && cfun->calls_alloca) |
2424 | var_end_seq = asan_emit_allocas_unpoison (virtual_stack_dynamic_rtx, |
2425 | virtual_stack_vars_rtx, |
2426 | var_end_seq); |
2427 | else if (hwasan_sanitize_allocas_p () && cfun->calls_alloca) |
2428 | /* When using out-of-line instrumentation we only want to emit one function |
2429 | call for clearing the tags in a region of shadow stack. When there are |
2430 | alloca calls in this frame we want to emit a call using the |
2431 | virtual_stack_dynamic_rtx, but when not we use the hwasan_frame_extent |
2432 | rtx we created in expand_stack_vars. */ |
2433 | var_end_seq = hwasan_emit_untag_frame (virtual_stack_dynamic_rtx, |
2434 | virtual_stack_vars_rtx); |
2435 | else if (hwasan_sanitize_stack_p ()) |
2436 | /* If no variables were stored on the stack, `hwasan_get_frame_extent` |
2437 | will return NULL_RTX and hence `hwasan_emit_untag_frame` will return |
2438 | NULL (i.e. an empty sequence). */ |
2439 | var_end_seq = hwasan_emit_untag_frame (hwasan_get_frame_extent (), |
2440 | virtual_stack_vars_rtx); |
2441 | |
2442 | fini_vars_expansion (); |
2443 | |
2444 | /* If there were any artificial non-ignored vars without rtl |
2445 | found earlier, see if deferred stack allocation hasn't assigned |
2446 | rtl to them. */ |
2447 | FOR_EACH_VEC_ELT_REVERSE (maybe_local_decls, i, var) |
2448 | { |
2449 | rtx rtl = DECL_RTL_IF_SET (var); |
2450 | |
2451 | /* Keep artificial non-ignored vars in cfun->local_decls |
2452 | chain until instantiate_decls. */ |
2453 | if (rtl && (MEM_P (rtl) || GET_CODE (rtl) == CONCAT)) |
2454 | add_local_decl (cfun, d: var); |
2455 | } |
2456 | |
2457 | /* If the target requires that FRAME_OFFSET be aligned, do it. */ |
2458 | if (STACK_ALIGNMENT_NEEDED) |
2459 | { |
2460 | HOST_WIDE_INT align = PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT; |
2461 | if (FRAME_GROWS_DOWNWARD) |
2462 | frame_offset = aligned_lower_bound (frame_offset, align); |
2463 | else |
2464 | frame_offset = aligned_upper_bound (frame_offset, align); |
2465 | } |
2466 | |
2467 | return var_end_seq; |
2468 | } |
2469 | |
2470 | |
2471 | /* If we need to produce a detailed dump, print the tree representation |
2472 | for STMT to the dump file. SINCE is the last RTX after which the RTL |
2473 | generated for STMT should have been appended. */ |
2474 | |
2475 | static void |
2476 | maybe_dump_rtl_for_gimple_stmt (gimple *stmt, rtx_insn *since) |
2477 | { |
2478 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2479 | { |
2480 | fprintf (stream: dump_file, format: "\n;; " ); |
2481 | print_gimple_stmt (dump_file, stmt, 0, |
2482 | TDF_SLIM | (dump_flags & TDF_LINENO)); |
2483 | fprintf (stream: dump_file, format: "\n" ); |
2484 | |
2485 | print_rtl (dump_file, since ? NEXT_INSN (insn: since) : since); |
2486 | } |
2487 | } |
2488 | |
2489 | /* Maps the blocks that do not contain tree labels to rtx labels. */ |
2490 | |
2491 | static hash_map<basic_block, rtx_code_label *> *lab_rtx_for_bb; |
2492 | |
2493 | /* Returns the label_rtx expression for a label starting basic block BB. */ |
2494 | |
2495 | static rtx_code_label * |
2496 | label_rtx_for_bb (basic_block bb ATTRIBUTE_UNUSED) |
2497 | { |
2498 | if (bb->flags & BB_RTL) |
2499 | return block_label (bb); |
2500 | |
2501 | rtx_code_label **elt = lab_rtx_for_bb->get (k: bb); |
2502 | if (elt) |
2503 | return *elt; |
2504 | |
2505 | /* Find the tree label if it is present. */ |
2506 | gimple_stmt_iterator gsi = gsi_start_bb (bb); |
2507 | glabel *lab_stmt; |
2508 | if (!gsi_end_p (i: gsi) |
2509 | && (lab_stmt = dyn_cast <glabel *> (p: gsi_stmt (i: gsi))) |
2510 | && !DECL_NONLOCAL (gimple_label_label (lab_stmt))) |
2511 | return jump_target_rtx (gimple_label_label (gs: lab_stmt)); |
2512 | |
2513 | rtx_code_label *l = gen_label_rtx (); |
2514 | lab_rtx_for_bb->put (k: bb, v: l); |
2515 | return l; |
2516 | } |
2517 | |
2518 | |
2519 | /* A subroutine of expand_gimple_cond. Given E, a fallthrough edge |
2520 | of a basic block where we just expanded the conditional at the end, |
2521 | possibly clean up the CFG and instruction sequence. LAST is the |
2522 | last instruction before the just emitted jump sequence. */ |
2523 | |
2524 | static void |
2525 | maybe_cleanup_end_of_block (edge e, rtx_insn *last) |
2526 | { |
2527 | /* Special case: when jumpif decides that the condition is |
2528 | trivial it emits an unconditional jump (and the necessary |
2529 | barrier). But we still have two edges, the fallthru one is |
2530 | wrong. purge_dead_edges would clean this up later. Unfortunately |
2531 | we have to insert insns (and split edges) before |
2532 | find_many_sub_basic_blocks and hence before purge_dead_edges. |
2533 | But splitting edges might create new blocks which depend on the |
2534 | fact that if there are two edges there's no barrier. So the |
2535 | barrier would get lost and verify_flow_info would ICE. Instead |
2536 | of auditing all edge splitters to care for the barrier (which |
2537 | normally isn't there in a cleaned CFG), fix it here. */ |
2538 | if (BARRIER_P (get_last_insn ())) |
2539 | { |
2540 | rtx_insn *insn; |
2541 | remove_edge (e); |
2542 | /* Now, we have a single successor block, if we have insns to |
2543 | insert on the remaining edge we potentially will insert |
2544 | it at the end of this block (if the dest block isn't feasible) |
2545 | in order to avoid splitting the edge. This insertion will take |
2546 | place in front of the last jump. But we might have emitted |
2547 | multiple jumps (conditional and one unconditional) to the |
2548 | same destination. Inserting in front of the last one then |
2549 | is a problem. See PR 40021. We fix this by deleting all |
2550 | jumps except the last unconditional one. */ |
2551 | insn = PREV_INSN (insn: get_last_insn ()); |
2552 | /* Make sure we have an unconditional jump. Otherwise we're |
2553 | confused. */ |
2554 | gcc_assert (JUMP_P (insn) && !any_condjump_p (insn)); |
2555 | for (insn = PREV_INSN (insn); insn != last;) |
2556 | { |
2557 | insn = PREV_INSN (insn); |
2558 | if (JUMP_P (NEXT_INSN (insn))) |
2559 | { |
2560 | if (!any_condjump_p (NEXT_INSN (insn))) |
2561 | { |
2562 | gcc_assert (BARRIER_P (NEXT_INSN (NEXT_INSN (insn)))); |
2563 | delete_insn (NEXT_INSN (insn: NEXT_INSN (insn))); |
2564 | } |
2565 | delete_insn (NEXT_INSN (insn)); |
2566 | } |
2567 | } |
2568 | } |
2569 | } |
2570 | |
2571 | /* A subroutine of expand_gimple_basic_block. Expand one GIMPLE_COND. |
2572 | Returns a new basic block if we've terminated the current basic |
2573 | block and created a new one. */ |
2574 | |
2575 | static basic_block |
2576 | expand_gimple_cond (basic_block bb, gcond *stmt) |
2577 | { |
2578 | basic_block new_bb, dest; |
2579 | edge true_edge; |
2580 | edge false_edge; |
2581 | rtx_insn *last2, *last; |
2582 | enum tree_code code; |
2583 | tree op0, op1; |
2584 | |
2585 | code = gimple_cond_code (gs: stmt); |
2586 | op0 = gimple_cond_lhs (gs: stmt); |
2587 | op1 = gimple_cond_rhs (gs: stmt); |
2588 | /* We're sometimes presented with such code: |
2589 | D.123_1 = x < y; |
2590 | if (D.123_1 != 0) |
2591 | ... |
2592 | This would expand to two comparisons which then later might |
2593 | be cleaned up by combine. But some pattern matchers like if-conversion |
2594 | work better when there's only one compare, so make up for this |
2595 | here as special exception if TER would have made the same change. */ |
2596 | if (SA.values |
2597 | && TREE_CODE (op0) == SSA_NAME |
2598 | && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE |
2599 | && TREE_CODE (op1) == INTEGER_CST |
2600 | && ((gimple_cond_code (gs: stmt) == NE_EXPR |
2601 | && integer_zerop (op1)) |
2602 | || (gimple_cond_code (gs: stmt) == EQ_EXPR |
2603 | && integer_onep (op1))) |
2604 | && bitmap_bit_p (SA.values, SSA_NAME_VERSION (op0))) |
2605 | { |
2606 | gimple *second = SSA_NAME_DEF_STMT (op0); |
2607 | if (gimple_code (g: second) == GIMPLE_ASSIGN) |
2608 | { |
2609 | enum tree_code code2 = gimple_assign_rhs_code (gs: second); |
2610 | if (TREE_CODE_CLASS (code2) == tcc_comparison) |
2611 | { |
2612 | code = code2; |
2613 | op0 = gimple_assign_rhs1 (gs: second); |
2614 | op1 = gimple_assign_rhs2 (gs: second); |
2615 | } |
2616 | /* If jumps are cheap and the target does not support conditional |
2617 | compare, turn some more codes into jumpy sequences. */ |
2618 | else if (BRANCH_COST (optimize_insn_for_speed_p (), false) < 4 |
2619 | && targetm.gen_ccmp_first == NULL) |
2620 | { |
2621 | if ((code2 == BIT_AND_EXPR |
2622 | && TYPE_PRECISION (TREE_TYPE (op0)) == 1 |
2623 | && TREE_CODE (gimple_assign_rhs2 (second)) != INTEGER_CST) |
2624 | || code2 == TRUTH_AND_EXPR) |
2625 | { |
2626 | code = TRUTH_ANDIF_EXPR; |
2627 | op0 = gimple_assign_rhs1 (gs: second); |
2628 | op1 = gimple_assign_rhs2 (gs: second); |
2629 | } |
2630 | else if (code2 == BIT_IOR_EXPR || code2 == TRUTH_OR_EXPR) |
2631 | { |
2632 | code = TRUTH_ORIF_EXPR; |
2633 | op0 = gimple_assign_rhs1 (gs: second); |
2634 | op1 = gimple_assign_rhs2 (gs: second); |
2635 | } |
2636 | } |
2637 | } |
2638 | } |
2639 | |
2640 | /* Optimize (x % C1) == C2 or (x % C1) != C2 if it is beneficial |
2641 | into (x - C2) * C3 < C4. */ |
2642 | if ((code == EQ_EXPR || code == NE_EXPR) |
2643 | && TREE_CODE (op0) == SSA_NAME |
2644 | && TREE_CODE (op1) == INTEGER_CST) |
2645 | code = maybe_optimize_mod_cmp (code, &op0, &op1); |
2646 | |
2647 | /* Optimize (x - y) < 0 into x < y if x - y has undefined overflow. */ |
2648 | if (!TYPE_UNSIGNED (TREE_TYPE (op0)) |
2649 | && (code == LT_EXPR || code == LE_EXPR |
2650 | || code == GT_EXPR || code == GE_EXPR) |
2651 | && integer_zerop (op1) |
2652 | && TREE_CODE (op0) == SSA_NAME) |
2653 | maybe_optimize_sub_cmp_0 (code, &op0, &op1); |
2654 | |
2655 | last2 = last = get_last_insn (); |
2656 | |
2657 | extract_true_false_edges_from_block (bb, &true_edge, &false_edge); |
2658 | set_curr_insn_location (gimple_location (g: stmt)); |
2659 | |
2660 | /* These flags have no purpose in RTL land. */ |
2661 | true_edge->flags &= ~EDGE_TRUE_VALUE; |
2662 | false_edge->flags &= ~EDGE_FALSE_VALUE; |
2663 | |
2664 | /* We can either have a pure conditional jump with one fallthru edge or |
2665 | two-way jump that needs to be decomposed into two basic blocks. */ |
2666 | if (false_edge->dest == bb->next_bb) |
2667 | { |
2668 | jumpif_1 (code, op0, op1, label_rtx_for_bb (bb: true_edge->dest), |
2669 | true_edge->probability); |
2670 | maybe_dump_rtl_for_gimple_stmt (stmt, since: last); |
2671 | if (true_edge->goto_locus != UNKNOWN_LOCATION) |
2672 | set_curr_insn_location (true_edge->goto_locus); |
2673 | false_edge->flags |= EDGE_FALLTHRU; |
2674 | maybe_cleanup_end_of_block (e: false_edge, last); |
2675 | return NULL; |
2676 | } |
2677 | if (true_edge->dest == bb->next_bb) |
2678 | { |
2679 | jumpifnot_1 (code, op0, op1, label_rtx_for_bb (bb: false_edge->dest), |
2680 | false_edge->probability); |
2681 | maybe_dump_rtl_for_gimple_stmt (stmt, since: last); |
2682 | if (false_edge->goto_locus != UNKNOWN_LOCATION) |
2683 | set_curr_insn_location (false_edge->goto_locus); |
2684 | true_edge->flags |= EDGE_FALLTHRU; |
2685 | maybe_cleanup_end_of_block (e: true_edge, last); |
2686 | return NULL; |
2687 | } |
2688 | |
2689 | jumpif_1 (code, op0, op1, label_rtx_for_bb (bb: true_edge->dest), |
2690 | true_edge->probability); |
2691 | last = get_last_insn (); |
2692 | if (false_edge->goto_locus != UNKNOWN_LOCATION) |
2693 | set_curr_insn_location (false_edge->goto_locus); |
2694 | emit_jump (label_rtx_for_bb (bb: false_edge->dest)); |
2695 | |
2696 | BB_END (bb) = last; |
2697 | if (BARRIER_P (BB_END (bb))) |
2698 | BB_END (bb) = PREV_INSN (BB_END (bb)); |
2699 | update_bb_for_insn (bb); |
2700 | |
2701 | new_bb = create_basic_block (NEXT_INSN (insn: last), get_last_insn (), bb); |
2702 | dest = false_edge->dest; |
2703 | redirect_edge_succ (false_edge, new_bb); |
2704 | false_edge->flags |= EDGE_FALLTHRU; |
2705 | new_bb->count = false_edge->count (); |
2706 | loop_p loop = find_common_loop (bb->loop_father, dest->loop_father); |
2707 | add_bb_to_loop (new_bb, loop); |
2708 | if (loop->latch == bb |
2709 | && loop->header == dest) |
2710 | loop->latch = new_bb; |
2711 | make_single_succ_edge (new_bb, dest, 0); |
2712 | if (BARRIER_P (BB_END (new_bb))) |
2713 | BB_END (new_bb) = PREV_INSN (BB_END (new_bb)); |
2714 | update_bb_for_insn (new_bb); |
2715 | |
2716 | maybe_dump_rtl_for_gimple_stmt (stmt, since: last2); |
2717 | |
2718 | if (true_edge->goto_locus != UNKNOWN_LOCATION) |
2719 | { |
2720 | set_curr_insn_location (true_edge->goto_locus); |
2721 | true_edge->goto_locus = curr_insn_location (); |
2722 | } |
2723 | |
2724 | return new_bb; |
2725 | } |
2726 | |
2727 | /* Mark all calls that can have a transaction restart. */ |
2728 | |
2729 | static void |
2730 | mark_transaction_restart_calls (gimple *stmt) |
2731 | { |
2732 | struct tm_restart_node dummy; |
2733 | tm_restart_node **slot; |
2734 | |
2735 | if (!cfun->gimple_df->tm_restart) |
2736 | return; |
2737 | |
2738 | dummy.stmt = stmt; |
2739 | slot = cfun->gimple_df->tm_restart->find_slot (value: &dummy, insert: NO_INSERT); |
2740 | if (slot) |
2741 | { |
2742 | struct tm_restart_node *n = *slot; |
2743 | tree list = n->label_or_list; |
2744 | rtx_insn *insn; |
2745 | |
2746 | for (insn = next_real_insn (get_last_insn ()); |
2747 | !CALL_P (insn); |
2748 | insn = next_real_insn (insn)) |
2749 | continue; |
2750 | |
2751 | if (TREE_CODE (list) == LABEL_DECL) |
2752 | add_reg_note (insn, REG_TM, label_rtx (list)); |
2753 | else |
2754 | for (; list ; list = TREE_CHAIN (list)) |
2755 | add_reg_note (insn, REG_TM, label_rtx (TREE_VALUE (list))); |
2756 | } |
2757 | } |
2758 | |
2759 | /* A subroutine of expand_gimple_stmt_1, expanding one GIMPLE_CALL |
2760 | statement STMT. */ |
2761 | |
2762 | static void |
2763 | expand_call_stmt (gcall *stmt) |
2764 | { |
2765 | tree exp, decl, lhs; |
2766 | bool builtin_p; |
2767 | size_t i; |
2768 | |
2769 | if (gimple_call_internal_p (gs: stmt)) |
2770 | { |
2771 | expand_internal_call (stmt); |
2772 | return; |
2773 | } |
2774 | |
2775 | /* If this is a call to a built-in function and it has no effect other |
2776 | than setting the lhs, try to implement it using an internal function |
2777 | instead. */ |
2778 | decl = gimple_call_fndecl (gs: stmt); |
2779 | if (gimple_call_lhs (gs: stmt) |
2780 | && !gimple_has_side_effects (stmt) |
2781 | && (optimize || (decl && called_as_built_in (decl)))) |
2782 | { |
2783 | internal_fn ifn = replacement_internal_fn (stmt); |
2784 | if (ifn != IFN_LAST) |
2785 | { |
2786 | expand_internal_call (ifn, stmt); |
2787 | return; |
2788 | } |
2789 | } |
2790 | |
2791 | exp = build_vl_exp (CALL_EXPR, gimple_call_num_args (gs: stmt) + 3); |
2792 | |
2793 | CALL_EXPR_FN (exp) = gimple_call_fn (gs: stmt); |
2794 | builtin_p = decl && fndecl_built_in_p (node: decl); |
2795 | |
2796 | /* If this is not a builtin function, the function type through which the |
2797 | call is made may be different from the type of the function. */ |
2798 | if (!builtin_p) |
2799 | CALL_EXPR_FN (exp) |
2800 | = fold_convert (build_pointer_type (gimple_call_fntype (stmt)), |
2801 | CALL_EXPR_FN (exp)); |
2802 | |
2803 | TREE_TYPE (exp) = gimple_call_return_type (gs: stmt); |
2804 | CALL_EXPR_STATIC_CHAIN (exp) = gimple_call_chain (gs: stmt); |
2805 | |
2806 | for (i = 0; i < gimple_call_num_args (gs: stmt); i++) |
2807 | { |
2808 | tree arg = gimple_call_arg (gs: stmt, index: i); |
2809 | gimple *def; |
2810 | /* TER addresses into arguments of builtin functions so we have a |
2811 | chance to infer more correct alignment information. See PR39954. */ |
2812 | if (builtin_p |
2813 | && TREE_CODE (arg) == SSA_NAME |
2814 | && (def = get_gimple_for_ssa_name (exp: arg)) |
2815 | && gimple_assign_rhs_code (gs: def) == ADDR_EXPR) |
2816 | arg = gimple_assign_rhs1 (gs: def); |
2817 | CALL_EXPR_ARG (exp, i) = arg; |
2818 | } |
2819 | |
2820 | if (gimple_has_side_effects (stmt) |
2821 | /* ??? Downstream in expand_expr_real_1 we assume that expressions |
2822 | w/o side-effects do not throw so work around this here. */ |
2823 | || stmt_could_throw_p (cfun, stmt)) |
2824 | TREE_SIDE_EFFECTS (exp) = 1; |
2825 | |
2826 | if (gimple_call_nothrow_p (s: stmt)) |
2827 | TREE_NOTHROW (exp) = 1; |
2828 | |
2829 | CALL_EXPR_TAILCALL (exp) = gimple_call_tail_p (s: stmt); |
2830 | CALL_EXPR_MUST_TAIL_CALL (exp) = gimple_call_must_tail_p (s: stmt); |
2831 | CALL_EXPR_RETURN_SLOT_OPT (exp) = gimple_call_return_slot_opt_p (s: stmt); |
2832 | if (decl |
2833 | && fndecl_built_in_p (node: decl, klass: BUILT_IN_NORMAL) |
2834 | && ALLOCA_FUNCTION_CODE_P (DECL_FUNCTION_CODE (decl))) |
2835 | CALL_ALLOCA_FOR_VAR_P (exp) = gimple_call_alloca_for_var_p (s: stmt); |
2836 | else |
2837 | CALL_FROM_THUNK_P (exp) = gimple_call_from_thunk_p (s: stmt); |
2838 | CALL_EXPR_VA_ARG_PACK (exp) = gimple_call_va_arg_pack_p (s: stmt); |
2839 | CALL_EXPR_BY_DESCRIPTOR (exp) = gimple_call_by_descriptor_p (s: stmt); |
2840 | SET_EXPR_LOCATION (exp, gimple_location (stmt)); |
2841 | |
2842 | /* Must come after copying location. */ |
2843 | copy_warning (exp, stmt); |
2844 | |
2845 | /* Ensure RTL is created for debug args. */ |
2846 | if (decl && DECL_HAS_DEBUG_ARGS_P (decl)) |
2847 | { |
2848 | vec<tree, va_gc> **debug_args = decl_debug_args_lookup (decl); |
2849 | unsigned int ix; |
2850 | tree dtemp; |
2851 | |
2852 | if (debug_args) |
2853 | for (ix = 1; (*debug_args)->iterate (ix, ptr: &dtemp); ix += 2) |
2854 | { |
2855 | gcc_assert (TREE_CODE (dtemp) == DEBUG_EXPR_DECL); |
2856 | expand_debug_expr (dtemp); |
2857 | } |
2858 | } |
2859 | |
2860 | rtx_insn *before_call = get_last_insn (); |
2861 | lhs = gimple_call_lhs (gs: stmt); |
2862 | if (lhs) |
2863 | expand_assignment (lhs, exp, false); |
2864 | else |
2865 | expand_expr (exp, const0_rtx, VOIDmode, modifier: EXPAND_NORMAL); |
2866 | |
2867 | /* If the gimple call is an indirect call and has 'nocf_check' |
2868 | attribute find a generated CALL insn to mark it as no |
2869 | control-flow verification is needed. */ |
2870 | if (gimple_call_nocf_check_p (gs: stmt) |
2871 | && !gimple_call_fndecl (gs: stmt)) |
2872 | { |
2873 | rtx_insn *last = get_last_insn (); |
2874 | while (!CALL_P (last) |
2875 | && last != before_call) |
2876 | last = PREV_INSN (insn: last); |
2877 | |
2878 | if (last != before_call) |
2879 | add_reg_note (last, REG_CALL_NOCF_CHECK, const0_rtx); |
2880 | } |
2881 | |
2882 | mark_transaction_restart_calls (stmt); |
2883 | } |
2884 | |
2885 | |
2886 | /* Generate RTL for an asm statement (explicit assembler code). |
2887 | STRING is a STRING_CST node containing the assembler code text, |
2888 | or an ADDR_EXPR containing a STRING_CST. VOL nonzero means the |
2889 | insn is volatile; don't optimize it. */ |
2890 | |
2891 | static void |
2892 | expand_asm_loc (tree string, int vol, location_t locus) |
2893 | { |
2894 | rtx body; |
2895 | |
2896 | body = gen_rtx_ASM_INPUT_loc (VOIDmode, |
2897 | ggc_strdup (TREE_STRING_POINTER (string)), |
2898 | locus); |
2899 | |
2900 | MEM_VOLATILE_P (body) = vol; |
2901 | |
2902 | /* Non-empty basic ASM implicitly clobbers memory. */ |
2903 | if (TREE_STRING_LENGTH (string) != 0) |
2904 | { |
2905 | rtx asm_op, clob; |
2906 | unsigned i, nclobbers; |
2907 | auto_vec<rtx> input_rvec, output_rvec; |
2908 | auto_vec<machine_mode> input_mode; |
2909 | auto_vec<const char *> constraints; |
2910 | auto_vec<rtx> use_rvec; |
2911 | auto_vec<rtx> clobber_rvec; |
2912 | HARD_REG_SET clobbered_regs; |
2913 | CLEAR_HARD_REG_SET (set&: clobbered_regs); |
2914 | |
2915 | clob = gen_rtx_MEM (BLKmode, gen_rtx_SCRATCH (VOIDmode)); |
2916 | clobber_rvec.safe_push (obj: clob); |
2917 | |
2918 | if (targetm.md_asm_adjust) |
2919 | targetm.md_asm_adjust (output_rvec, input_rvec, input_mode, |
2920 | constraints, use_rvec, clobber_rvec, |
2921 | clobbered_regs, locus); |
2922 | |
2923 | asm_op = body; |
2924 | nclobbers = clobber_rvec.length (); |
2925 | auto nuses = use_rvec.length (); |
2926 | body = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (1 + nuses + nclobbers)); |
2927 | |
2928 | i = 0; |
2929 | XVECEXP (body, 0, i++) = asm_op; |
2930 | for (rtx use : use_rvec) |
2931 | XVECEXP (body, 0, i++) = gen_rtx_USE (VOIDmode, use); |
2932 | for (rtx clobber : clobber_rvec) |
2933 | XVECEXP (body, 0, i++) = gen_rtx_CLOBBER (VOIDmode, clobber); |
2934 | } |
2935 | |
2936 | emit_insn (body); |
2937 | } |
2938 | |
2939 | /* Return the number of times character C occurs in string S. */ |
2940 | static int |
2941 | n_occurrences (int c, const char *s) |
2942 | { |
2943 | int n = 0; |
2944 | while (*s) |
2945 | n += (*s++ == c); |
2946 | return n; |
2947 | } |
2948 | |
2949 | /* A subroutine of expand_asm_operands. Check that all operands have |
2950 | the same number of alternatives. Return true if so. */ |
2951 | |
2952 | static bool |
2953 | check_operand_nalternatives (const vec<const char *> &constraints) |
2954 | { |
2955 | unsigned len = constraints.length(); |
2956 | if (len > 0) |
2957 | { |
2958 | int nalternatives = n_occurrences (c: ',', s: constraints[0]); |
2959 | |
2960 | if (nalternatives + 1 > MAX_RECOG_ALTERNATIVES) |
2961 | { |
2962 | error ("too many alternatives in %<asm%>" ); |
2963 | return false; |
2964 | } |
2965 | |
2966 | for (unsigned i = 1; i < len; ++i) |
2967 | if (n_occurrences (c: ',', s: constraints[i]) != nalternatives) |
2968 | { |
2969 | error ("operand constraints for %<asm%> differ " |
2970 | "in number of alternatives" ); |
2971 | return false; |
2972 | } |
2973 | } |
2974 | return true; |
2975 | } |
2976 | |
2977 | /* Check for overlap between registers marked in CLOBBERED_REGS and |
2978 | anything inappropriate in T. Emit error and return the register |
2979 | variable definition for error, NULL_TREE for ok. */ |
2980 | |
2981 | static bool |
2982 | tree_conflicts_with_clobbers_p (tree t, HARD_REG_SET *clobbered_regs, |
2983 | location_t loc) |
2984 | { |
2985 | /* Conflicts between asm-declared register variables and the clobber |
2986 | list are not allowed. */ |
2987 | tree overlap = tree_overlaps_hard_reg_set (t, clobbered_regs); |
2988 | |
2989 | if (overlap) |
2990 | { |
2991 | error_at (loc, "%<asm%> specifier for variable %qE conflicts with " |
2992 | "%<asm%> clobber list" , DECL_NAME (overlap)); |
2993 | |
2994 | /* Reset registerness to stop multiple errors emitted for a single |
2995 | variable. */ |
2996 | DECL_REGISTER (overlap) = 0; |
2997 | return true; |
2998 | } |
2999 | |
3000 | return false; |
3001 | } |
3002 | |
3003 | /* Check that the given REGNO spanning NREGS is a valid |
3004 | asm clobber operand. Some HW registers cannot be |
3005 | saved/restored, hence they should not be clobbered by |
3006 | asm statements. */ |
3007 | static bool |
3008 | asm_clobber_reg_is_valid (int regno, int nregs, const char *regname) |
3009 | { |
3010 | bool is_valid = true; |
3011 | HARD_REG_SET regset; |
3012 | |
3013 | CLEAR_HARD_REG_SET (set&: regset); |
3014 | |
3015 | add_range_to_hard_reg_set (regs: ®set, regno, nregs); |
3016 | |
3017 | /* Clobbering the PIC register is an error. */ |
3018 | if (PIC_OFFSET_TABLE_REGNUM != INVALID_REGNUM |
3019 | && overlaps_hard_reg_set_p (regs: regset, Pmode, PIC_OFFSET_TABLE_REGNUM)) |
3020 | { |
3021 | /* ??? Diagnose during gimplification? */ |
3022 | error ("PIC register clobbered by %qs in %<asm%>" , regname); |
3023 | is_valid = false; |
3024 | } |
3025 | else if (!in_hard_reg_set_p |
3026 | (accessible_reg_set, reg_raw_mode[regno], regno)) |
3027 | { |
3028 | /* ??? Diagnose during gimplification? */ |
3029 | error ("the register %qs cannot be clobbered in %<asm%>" |
3030 | " for the current target" , regname); |
3031 | is_valid = false; |
3032 | } |
3033 | |
3034 | /* Clobbering the stack pointer register is deprecated. GCC expects |
3035 | the value of the stack pointer after an asm statement to be the same |
3036 | as it was before, so no asm can validly clobber the stack pointer in |
3037 | the usual sense. Adding the stack pointer to the clobber list has |
3038 | traditionally had some undocumented and somewhat obscure side-effects. */ |
3039 | if (overlaps_hard_reg_set_p (regs: regset, Pmode, STACK_POINTER_REGNUM)) |
3040 | { |
3041 | crtl->sp_is_clobbered_by_asm = true; |
3042 | if (warning (OPT_Wdeprecated, "listing the stack pointer register" |
3043 | " %qs in a clobber list is deprecated" , regname)) |
3044 | inform (input_location, "the value of the stack pointer after" |
3045 | " an %<asm%> statement must be the same as it was before" |
3046 | " the statement" ); |
3047 | } |
3048 | |
3049 | return is_valid; |
3050 | } |
3051 | |
3052 | /* Generate RTL for an asm statement with arguments. |
3053 | STRING is the instruction template. |
3054 | OUTPUTS is a list of output arguments (lvalues); INPUTS a list of inputs. |
3055 | Each output or input has an expression in the TREE_VALUE and |
3056 | a tree list in TREE_PURPOSE which in turn contains a constraint |
3057 | name in TREE_VALUE (or NULL_TREE) and a constraint string |
3058 | in TREE_PURPOSE. |
3059 | CLOBBERS is a list of STRING_CST nodes each naming a hard register |
3060 | that is clobbered by this insn. |
3061 | |
3062 | LABELS is a list of labels, and if LABELS is non-NULL, FALLTHRU_BB |
3063 | should be the fallthru basic block of the asm goto. |
3064 | |
3065 | Not all kinds of lvalue that may appear in OUTPUTS can be stored directly. |
3066 | Some elements of OUTPUTS may be replaced with trees representing temporary |
3067 | values. The caller should copy those temporary values to the originally |
3068 | specified lvalues. |
3069 | |
3070 | VOL nonzero means the insn is volatile; don't optimize it. */ |
3071 | |
3072 | static void |
3073 | expand_asm_stmt (gasm *stmt) |
3074 | { |
3075 | class save_input_location |
3076 | { |
3077 | location_t old; |
3078 | |
3079 | public: |
3080 | explicit save_input_location(location_t where) |
3081 | { |
3082 | old = input_location; |
3083 | input_location = where; |
3084 | } |
3085 | |
3086 | ~save_input_location() |
3087 | { |
3088 | input_location = old; |
3089 | } |
3090 | }; |
3091 | |
3092 | location_t locus = gimple_location (g: stmt); |
3093 | |
3094 | if (gimple_asm_input_p (asm_stmt: stmt)) |
3095 | { |
3096 | const char *s = gimple_asm_string (asm_stmt: stmt); |
3097 | tree string = build_string (strlen (s: s), s); |
3098 | expand_asm_loc (string, vol: gimple_asm_volatile_p (asm_stmt: stmt), locus); |
3099 | return; |
3100 | } |
3101 | |
3102 | /* There are some legacy diagnostics in here. */ |
3103 | save_input_location s_i_l(locus); |
3104 | |
3105 | unsigned noutputs = gimple_asm_noutputs (asm_stmt: stmt); |
3106 | unsigned ninputs = gimple_asm_ninputs (asm_stmt: stmt); |
3107 | unsigned nlabels = gimple_asm_nlabels (asm_stmt: stmt); |
3108 | unsigned i; |
3109 | bool error_seen = false; |
3110 | |
3111 | /* ??? Diagnose during gimplification? */ |
3112 | if (ninputs + noutputs + nlabels > MAX_RECOG_OPERANDS) |
3113 | { |
3114 | error_at (locus, "more than %d operands in %<asm%>" , MAX_RECOG_OPERANDS); |
3115 | return; |
3116 | } |
3117 | |
3118 | auto_vec<tree, MAX_RECOG_OPERANDS> output_tvec; |
3119 | auto_vec<tree, MAX_RECOG_OPERANDS> input_tvec; |
3120 | auto_vec<const char *, MAX_RECOG_OPERANDS> constraints; |
3121 | |
3122 | /* Copy the gimple vectors into new vectors that we can manipulate. */ |
3123 | |
3124 | output_tvec.safe_grow (len: noutputs, exact: true); |
3125 | input_tvec.safe_grow (len: ninputs, exact: true); |
3126 | constraints.safe_grow (len: noutputs + ninputs, exact: true); |
3127 | |
3128 | for (i = 0; i < noutputs; ++i) |
3129 | { |
3130 | tree t = gimple_asm_output_op (asm_stmt: stmt, index: i); |
3131 | output_tvec[i] = TREE_VALUE (t); |
3132 | constraints[i] = TREE_STRING_POINTER (TREE_VALUE (TREE_PURPOSE (t))); |
3133 | } |
3134 | for (i = 0; i < ninputs; i++) |
3135 | { |
3136 | tree t = gimple_asm_input_op (asm_stmt: stmt, index: i); |
3137 | input_tvec[i] = TREE_VALUE (t); |
3138 | constraints[i + noutputs] |
3139 | = TREE_STRING_POINTER (TREE_VALUE (TREE_PURPOSE (t))); |
3140 | } |
3141 | |
3142 | /* ??? Diagnose during gimplification? */ |
3143 | if (! check_operand_nalternatives (constraints)) |
3144 | return; |
3145 | |
3146 | /* Count the number of meaningful clobbered registers, ignoring what |
3147 | we would ignore later. */ |
3148 | auto_vec<rtx> clobber_rvec; |
3149 | HARD_REG_SET clobbered_regs; |
3150 | CLEAR_HARD_REG_SET (set&: clobbered_regs); |
3151 | |
3152 | if (unsigned n = gimple_asm_nclobbers (asm_stmt: stmt)) |
3153 | { |
3154 | clobber_rvec.reserve (nelems: n); |
3155 | for (i = 0; i < n; i++) |
3156 | { |
3157 | tree t = gimple_asm_clobber_op (asm_stmt: stmt, index: i); |
3158 | const char *regname = TREE_STRING_POINTER (TREE_VALUE (t)); |
3159 | int nregs, j; |
3160 | |
3161 | j = decode_reg_name_and_count (regname, &nregs); |
3162 | if (j < 0) |
3163 | { |
3164 | if (j == -2) |
3165 | { |
3166 | /* ??? Diagnose during gimplification? */ |
3167 | error_at (locus, "unknown register name %qs in %<asm%>" , |
3168 | regname); |
3169 | error_seen = true; |
3170 | } |
3171 | else if (j == -4) |
3172 | { |
3173 | rtx x = gen_rtx_MEM (BLKmode, gen_rtx_SCRATCH (VOIDmode)); |
3174 | clobber_rvec.safe_push (obj: x); |
3175 | } |
3176 | else |
3177 | { |
3178 | /* Otherwise we should have -1 == empty string |
3179 | or -3 == cc, which is not a register. */ |
3180 | gcc_assert (j == -1 || j == -3); |
3181 | } |
3182 | } |
3183 | else |
3184 | for (int reg = j; reg < j + nregs; reg++) |
3185 | { |
3186 | if (!asm_clobber_reg_is_valid (regno: reg, nregs, regname)) |
3187 | return; |
3188 | |
3189 | SET_HARD_REG_BIT (set&: clobbered_regs, bit: reg); |
3190 | rtx x = gen_rtx_REG (reg_raw_mode[reg], reg); |
3191 | clobber_rvec.safe_push (obj: x); |
3192 | } |
3193 | } |
3194 | } |
3195 | |
3196 | /* First pass over inputs and outputs checks validity and sets |
3197 | mark_addressable if needed. */ |
3198 | /* ??? Diagnose during gimplification? */ |
3199 | |
3200 | for (i = 0; i < noutputs; ++i) |
3201 | { |
3202 | tree val = output_tvec[i]; |
3203 | tree type = TREE_TYPE (val); |
3204 | const char *constraint; |
3205 | bool is_inout; |
3206 | bool allows_reg; |
3207 | bool allows_mem; |
3208 | |
3209 | /* Try to parse the output constraint. If that fails, there's |
3210 | no point in going further. */ |
3211 | constraint = constraints[i]; |
3212 | if (!parse_output_constraint (&constraint, i, ninputs, noutputs, |
3213 | &allows_mem, &allows_reg, &is_inout)) |
3214 | return; |
3215 | |
3216 | /* If the output is a hard register, verify it doesn't conflict with |
3217 | any other operand's possible hard register use. */ |
3218 | if (DECL_P (val) |
3219 | && REG_P (DECL_RTL (val)) |
3220 | && HARD_REGISTER_P (DECL_RTL (val))) |
3221 | { |
3222 | unsigned j, output_hregno = REGNO (DECL_RTL (val)); |
3223 | bool early_clobber_p = strchr (s: constraints[i], c: '&') != NULL; |
3224 | unsigned long match; |
3225 | |
3226 | /* Verify the other outputs do not use the same hard register. */ |
3227 | for (j = i + 1; j < noutputs; ++j) |
3228 | if (DECL_P (output_tvec[j]) |
3229 | && REG_P (DECL_RTL (output_tvec[j])) |
3230 | && HARD_REGISTER_P (DECL_RTL (output_tvec[j])) |
3231 | && output_hregno == REGNO (DECL_RTL (output_tvec[j]))) |
3232 | { |
3233 | error_at (locus, "invalid hard register usage between output " |
3234 | "operands" ); |
3235 | error_seen = true; |
3236 | } |
3237 | |
3238 | /* Verify matching constraint operands use the same hard register |
3239 | and that the non-matching constraint operands do not use the same |
3240 | hard register if the output is an early clobber operand. */ |
3241 | for (j = 0; j < ninputs; ++j) |
3242 | if (DECL_P (input_tvec[j]) |
3243 | && REG_P (DECL_RTL (input_tvec[j])) |
3244 | && HARD_REGISTER_P (DECL_RTL (input_tvec[j]))) |
3245 | { |
3246 | unsigned input_hregno = REGNO (DECL_RTL (input_tvec[j])); |
3247 | switch (*constraints[j + noutputs]) |
3248 | { |
3249 | case '0': case '1': case '2': case '3': case '4': |
3250 | case '5': case '6': case '7': case '8': case '9': |
3251 | match = strtoul (nptr: constraints[j + noutputs], NULL, base: 10); |
3252 | break; |
3253 | default: |
3254 | match = ULONG_MAX; |
3255 | break; |
3256 | } |
3257 | if (i == match |
3258 | && output_hregno != input_hregno) |
3259 | { |
3260 | error_at (locus, "invalid hard register usage between " |
3261 | "output operand and matching constraint operand" ); |
3262 | error_seen = true; |
3263 | } |
3264 | else if (early_clobber_p |
3265 | && i != match |
3266 | && output_hregno == input_hregno) |
3267 | { |
3268 | error_at (locus, "invalid hard register usage between " |
3269 | "earlyclobber operand and input operand" ); |
3270 | error_seen = true; |
3271 | } |
3272 | } |
3273 | } |
3274 | |
3275 | if (! allows_reg |
3276 | && (allows_mem |
3277 | || is_inout |
3278 | || (DECL_P (val) |
3279 | && REG_P (DECL_RTL (val)) |
3280 | && GET_MODE (DECL_RTL (val)) != TYPE_MODE (type)))) |
3281 | mark_addressable (val); |
3282 | } |
3283 | |
3284 | for (i = 0; i < ninputs; ++i) |
3285 | { |
3286 | bool allows_reg, allows_mem; |
3287 | const char *constraint; |
3288 | |
3289 | constraint = constraints[i + noutputs]; |
3290 | if (! parse_input_constraint (&constraint, i, ninputs, noutputs, 0, |
3291 | constraints.address (), |
3292 | &allows_mem, &allows_reg)) |
3293 | return; |
3294 | |
3295 | if (! allows_reg && allows_mem) |
3296 | mark_addressable (input_tvec[i]); |
3297 | } |
3298 | |
3299 | /* Second pass evaluates arguments. */ |
3300 | |
3301 | /* Make sure stack is consistent for asm goto. */ |
3302 | if (nlabels > 0) |
3303 | do_pending_stack_adjust (); |
3304 | int old_generating_concat_p = generating_concat_p; |
3305 | |
3306 | /* Vector of RTX's of evaluated output operands. */ |
3307 | auto_vec<rtx, MAX_RECOG_OPERANDS> output_rvec; |
3308 | auto_vec<int, MAX_RECOG_OPERANDS> inout_opnum; |
3309 | rtx_insn *after_rtl_seq = NULL, *after_rtl_end = NULL; |
3310 | |
3311 | output_rvec.safe_grow (len: noutputs, exact: true); |
3312 | |
3313 | for (i = 0; i < noutputs; ++i) |
3314 | { |
3315 | tree val = output_tvec[i]; |
3316 | tree type = TREE_TYPE (val); |
3317 | bool is_inout, allows_reg, allows_mem, ok; |
3318 | rtx op; |
3319 | |
3320 | ok = parse_output_constraint (&constraints[i], i, ninputs, |
3321 | noutputs, &allows_mem, &allows_reg, |
3322 | &is_inout); |
3323 | gcc_assert (ok); |
3324 | |
3325 | /* If an output operand is not a decl or indirect ref and our constraint |
3326 | allows a register, make a temporary to act as an intermediate. |
3327 | Make the asm insn write into that, then we will copy it to |
3328 | the real output operand. Likewise for promoted variables. */ |
3329 | |
3330 | generating_concat_p = 0; |
3331 | |
3332 | gcc_assert (TREE_CODE (val) != INDIRECT_REF); |
3333 | if (((TREE_CODE (val) == MEM_REF |
3334 | && TREE_CODE (TREE_OPERAND (val, 0)) != ADDR_EXPR) |
3335 | && allows_mem) |
3336 | || (DECL_P (val) |
3337 | && (allows_mem || REG_P (DECL_RTL (val))) |
3338 | && ! (REG_P (DECL_RTL (val)) |
3339 | && GET_MODE (DECL_RTL (val)) != TYPE_MODE (type))) |
3340 | || ! allows_reg |
3341 | || is_inout |
3342 | || TREE_ADDRESSABLE (type) |
3343 | || (!tree_fits_poly_int64_p (TYPE_SIZE (type)) |
3344 | && !known_size_p (a: max_int_size_in_bytes (type)))) |
3345 | { |
3346 | op = expand_expr (exp: val, NULL_RTX, VOIDmode, |
3347 | modifier: !allows_reg ? EXPAND_MEMORY : EXPAND_WRITE); |
3348 | if (MEM_P (op)) |
3349 | op = validize_mem (op); |
3350 | |
3351 | if (! allows_reg && !MEM_P (op)) |
3352 | { |
3353 | error_at (locus, "output number %d not directly addressable" , i); |
3354 | error_seen = true; |
3355 | } |
3356 | if ((! allows_mem && MEM_P (op) && GET_MODE (op) != BLKmode) |
3357 | || GET_CODE (op) == CONCAT) |
3358 | { |
3359 | rtx old_op = op; |
3360 | op = gen_reg_rtx (GET_MODE (op)); |
3361 | |
3362 | generating_concat_p = old_generating_concat_p; |
3363 | |
3364 | if (is_inout) |
3365 | emit_move_insn (op, old_op); |
3366 | |
3367 | push_to_sequence2 (after_rtl_seq, after_rtl_end); |
3368 | emit_move_insn (old_op, op); |
3369 | after_rtl_seq = get_insns (); |
3370 | after_rtl_end = get_last_insn (); |
3371 | end_sequence (); |
3372 | } |
3373 | } |
3374 | else |
3375 | { |
3376 | op = assign_temp (type, 0, 1); |
3377 | op = validize_mem (op); |
3378 | if (!MEM_P (op) && TREE_CODE (val) == SSA_NAME) |
3379 | set_reg_attrs_for_decl_rtl (SSA_NAME_VAR (val), x: op); |
3380 | |
3381 | generating_concat_p = old_generating_concat_p; |
3382 | |
3383 | push_to_sequence2 (after_rtl_seq, after_rtl_end); |
3384 | expand_assignment (val, make_tree (type, op), false); |
3385 | after_rtl_seq = get_insns (); |
3386 | after_rtl_end = get_last_insn (); |
3387 | end_sequence (); |
3388 | } |
3389 | output_rvec[i] = op; |
3390 | |
3391 | if (is_inout) |
3392 | inout_opnum.safe_push (obj: i); |
3393 | } |
3394 | |
3395 | const char *str = gimple_asm_string (asm_stmt: stmt); |
3396 | if (error_seen) |
3397 | { |
3398 | ninputs = 0; |
3399 | noutputs = 0; |
3400 | inout_opnum.truncate (size: 0); |
3401 | output_rvec.truncate (size: 0); |
3402 | clobber_rvec.truncate (size: 0); |
3403 | constraints.truncate (size: 0); |
3404 | CLEAR_HARD_REG_SET (set&: clobbered_regs); |
3405 | str = "" ; |
3406 | } |
3407 | |
3408 | auto_vec<rtx, MAX_RECOG_OPERANDS> input_rvec; |
3409 | auto_vec<machine_mode, MAX_RECOG_OPERANDS> input_mode; |
3410 | |
3411 | input_rvec.safe_grow (len: ninputs, exact: true); |
3412 | input_mode.safe_grow (len: ninputs, exact: true); |
3413 | |
3414 | generating_concat_p = 0; |
3415 | |
3416 | for (i = 0; i < ninputs; ++i) |
3417 | { |
3418 | tree val = input_tvec[i]; |
3419 | tree type = TREE_TYPE (val); |
3420 | bool allows_reg, allows_mem, ok; |
3421 | const char *constraint; |
3422 | rtx op; |
3423 | |
3424 | constraint = constraints[i + noutputs]; |
3425 | ok = parse_input_constraint (&constraint, i, ninputs, noutputs, 0, |
3426 | constraints.address (), |
3427 | &allows_mem, &allows_reg); |
3428 | gcc_assert (ok); |
3429 | |
3430 | /* EXPAND_INITIALIZER will not generate code for valid initializer |
3431 | constants, but will still generate code for other types of operand. |
3432 | This is the behavior we want for constant constraints. */ |
3433 | op = expand_expr (exp: val, NULL_RTX, VOIDmode, |
3434 | modifier: allows_reg ? EXPAND_NORMAL |
3435 | : allows_mem ? EXPAND_MEMORY |
3436 | : EXPAND_INITIALIZER); |
3437 | |
3438 | /* Never pass a CONCAT to an ASM. */ |
3439 | if (GET_CODE (op) == CONCAT) |
3440 | op = force_reg (GET_MODE (op), op); |
3441 | else if (MEM_P (op)) |
3442 | op = validize_mem (op); |
3443 | |
3444 | if (asm_operand_ok (op, constraint, NULL) <= 0) |
3445 | { |
3446 | if (allows_reg && TYPE_MODE (type) != BLKmode) |
3447 | op = force_reg (TYPE_MODE (type), op); |
3448 | else if (!allows_mem) |
3449 | warning_at (locus, 0, "%<asm%> operand %d probably does not match " |
3450 | "constraints" , i + noutputs); |
3451 | else if (MEM_P (op)) |
3452 | { |
3453 | /* We won't recognize either volatile memory or memory |
3454 | with a queued address as available a memory_operand |
3455 | at this point. Ignore it: clearly this *is* a memory. */ |
3456 | } |
3457 | else |
3458 | gcc_unreachable (); |
3459 | } |
3460 | input_rvec[i] = op; |
3461 | input_mode[i] = TYPE_MODE (type); |
3462 | } |
3463 | |
3464 | /* For in-out operands, copy output rtx to input rtx. */ |
3465 | unsigned ninout = inout_opnum.length (); |
3466 | for (i = 0; i < ninout; i++) |
3467 | { |
3468 | int j = inout_opnum[i]; |
3469 | rtx o = output_rvec[j]; |
3470 | |
3471 | input_rvec.safe_push (obj: o); |
3472 | input_mode.safe_push (GET_MODE (o)); |
3473 | |
3474 | char buffer[16]; |
3475 | sprintf (s: buffer, format: "%d" , j); |
3476 | constraints.safe_push (ggc_strdup (buffer)); |
3477 | } |
3478 | ninputs += ninout; |
3479 | |
3480 | /* Sometimes we wish to automatically clobber registers across an asm. |
3481 | Case in point is when the i386 backend moved from cc0 to a hard reg -- |
3482 | maintaining source-level compatibility means automatically clobbering |
3483 | the flags register. */ |
3484 | rtx_insn *after_md_seq = NULL; |
3485 | auto_vec<rtx> use_rvec; |
3486 | if (targetm.md_asm_adjust) |
3487 | after_md_seq |
3488 | = targetm.md_asm_adjust (output_rvec, input_rvec, input_mode, |
3489 | constraints, use_rvec, clobber_rvec, |
3490 | clobbered_regs, locus); |
3491 | |
3492 | /* Do not allow the hook to change the output and input count, |
3493 | lest it mess up the operand numbering. */ |
3494 | gcc_assert (output_rvec.length() == noutputs); |
3495 | gcc_assert (input_rvec.length() == ninputs); |
3496 | gcc_assert (constraints.length() == noutputs + ninputs); |
3497 | |
3498 | /* But it certainly can adjust the uses and clobbers. */ |
3499 | unsigned nuses = use_rvec.length (); |
3500 | unsigned nclobbers = clobber_rvec.length (); |
3501 | |
3502 | /* Third pass checks for easy conflicts. */ |
3503 | /* ??? Why are we doing this on trees instead of rtx. */ |
3504 | |
3505 | bool clobber_conflict_found = 0; |
3506 | for (i = 0; i < noutputs; ++i) |
3507 | if (tree_conflicts_with_clobbers_p (t: output_tvec[i], clobbered_regs: &clobbered_regs, loc: locus)) |
3508 | clobber_conflict_found = 1; |
3509 | for (i = 0; i < ninputs - ninout; ++i) |
3510 | if (tree_conflicts_with_clobbers_p (t: input_tvec[i], clobbered_regs: &clobbered_regs, loc: locus)) |
3511 | clobber_conflict_found = 1; |
3512 | |
3513 | /* Make vectors for the expression-rtx, constraint strings, |
3514 | and named operands. */ |
3515 | |
3516 | rtvec argvec = rtvec_alloc (ninputs); |
3517 | rtvec constraintvec = rtvec_alloc (ninputs); |
3518 | rtvec labelvec = rtvec_alloc (nlabels); |
3519 | |
3520 | rtx body = gen_rtx_ASM_OPERANDS ((noutputs == 0 ? VOIDmode |
3521 | : GET_MODE (output_rvec[0])), |
3522 | ggc_strdup (str), |
3523 | "" , 0, argvec, constraintvec, |
3524 | labelvec, locus); |
3525 | MEM_VOLATILE_P (body) = gimple_asm_volatile_p (asm_stmt: stmt); |
3526 | |
3527 | for (i = 0; i < ninputs; ++i) |
3528 | { |
3529 | ASM_OPERANDS_INPUT (body, i) = input_rvec[i]; |
3530 | ASM_OPERANDS_INPUT_CONSTRAINT_EXP (body, i) |
3531 | = gen_rtx_ASM_INPUT_loc (input_mode[i], |
3532 | constraints[i + noutputs], |
3533 | locus); |
3534 | } |
3535 | |
3536 | /* Copy labels to the vector. */ |
3537 | rtx_code_label *fallthru_label = NULL; |
3538 | if (nlabels > 0) |
3539 | { |
3540 | basic_block fallthru_bb = NULL; |
3541 | edge fallthru = find_fallthru_edge (edges: gimple_bb (g: stmt)->succs); |
3542 | if (fallthru) |
3543 | fallthru_bb = fallthru->dest; |
3544 | |
3545 | for (i = 0; i < nlabels; ++i) |
3546 | { |
3547 | tree label = TREE_VALUE (gimple_asm_label_op (stmt, i)); |
3548 | rtx_insn *r; |
3549 | /* If asm goto has any labels in the fallthru basic block, use |
3550 | a label that we emit immediately after the asm goto. Expansion |
3551 | may insert further instructions into the same basic block after |
3552 | asm goto and if we don't do this, insertion of instructions on |
3553 | the fallthru edge might misbehave. See PR58670. */ |
3554 | if (fallthru_bb && label_to_block (cfun, label) == fallthru_bb) |
3555 | { |
3556 | if (fallthru_label == NULL_RTX) |
3557 | fallthru_label = gen_label_rtx (); |
3558 | r = fallthru_label; |
3559 | } |
3560 | else |
3561 | r = label_rtx (label); |
3562 | ASM_OPERANDS_LABEL (body, i) = gen_rtx_LABEL_REF (Pmode, r); |
3563 | } |
3564 | } |
3565 | |
3566 | /* Now, for each output, construct an rtx |
3567 | (set OUTPUT (asm_operands INSN OUTPUTCONSTRAINT OUTPUTNUMBER |
3568 | ARGVEC CONSTRAINTS OPNAMES)) |
3569 | If there is more than one, put them inside a PARALLEL. */ |
3570 | |
3571 | if (noutputs == 0 && nuses == 0 && nclobbers == 0) |
3572 | { |
3573 | /* No output operands: put in a raw ASM_OPERANDS rtx. */ |
3574 | if (nlabels > 0) |
3575 | emit_jump_insn (body); |
3576 | else |
3577 | emit_insn (body); |
3578 | } |
3579 | else if (noutputs == 1 && nuses == 0 && nclobbers == 0) |
3580 | { |
3581 | ASM_OPERANDS_OUTPUT_CONSTRAINT (body) = constraints[0]; |
3582 | if (nlabels > 0) |
3583 | emit_jump_insn (gen_rtx_SET (output_rvec[0], body)); |
3584 | else |
3585 | emit_insn (gen_rtx_SET (output_rvec[0], body)); |
3586 | } |
3587 | else |
3588 | { |
3589 | rtx obody = body; |
3590 | int num = noutputs; |
3591 | |
3592 | if (num == 0) |
3593 | num = 1; |
3594 | |
3595 | body = gen_rtx_PARALLEL (VOIDmode, |
3596 | rtvec_alloc (num + nuses + nclobbers)); |
3597 | |
3598 | /* For each output operand, store a SET. */ |
3599 | for (i = 0; i < noutputs; ++i) |
3600 | { |
3601 | rtx src, o = output_rvec[i]; |
3602 | if (i == 0) |
3603 | { |
3604 | ASM_OPERANDS_OUTPUT_CONSTRAINT (obody) = constraints[0]; |
3605 | src = obody; |
3606 | } |
3607 | else |
3608 | { |
3609 | src = gen_rtx_ASM_OPERANDS (GET_MODE (o), |
3610 | ASM_OPERANDS_TEMPLATE (obody), |
3611 | constraints[i], i, argvec, |
3612 | constraintvec, labelvec, locus); |
3613 | MEM_VOLATILE_P (src) = gimple_asm_volatile_p (asm_stmt: stmt); |
3614 | } |
3615 | XVECEXP (body, 0, i) = gen_rtx_SET (o, src); |
3616 | } |
3617 | |
3618 | /* If there are no outputs (but there are some clobbers) |
3619 | store the bare ASM_OPERANDS into the PARALLEL. */ |
3620 | if (i == 0) |
3621 | XVECEXP (body, 0, i++) = obody; |
3622 | |
3623 | /* Add the uses specified by the target hook. No checking should |
3624 | be needed since this doesn't come directly from user code. */ |
3625 | for (rtx use : use_rvec) |
3626 | XVECEXP (body, 0, i++) = gen_rtx_USE (VOIDmode, use); |
3627 | |
3628 | /* Store (clobber REG) for each clobbered register specified. */ |
3629 | for (unsigned j = 0; j < nclobbers; ++j) |
3630 | { |
3631 | rtx clobbered_reg = clobber_rvec[j]; |
3632 | |
3633 | /* Do sanity check for overlap between clobbers and respectively |
3634 | input and outputs that hasn't been handled. Such overlap |
3635 | should have been detected and reported above. */ |
3636 | if (!clobber_conflict_found && REG_P (clobbered_reg)) |
3637 | { |
3638 | /* We test the old body (obody) contents to avoid |
3639 | tripping over the under-construction body. */ |
3640 | for (unsigned k = 0; k < noutputs; ++k) |
3641 | if (reg_overlap_mentioned_p (clobbered_reg, output_rvec[k])) |
3642 | internal_error ("%<asm%> clobber conflict with " |
3643 | "output operand" ); |
3644 | |
3645 | for (unsigned k = 0; k < ninputs - ninout; ++k) |
3646 | if (reg_overlap_mentioned_p (clobbered_reg, input_rvec[k])) |
3647 | internal_error ("%<asm%> clobber conflict with " |
3648 | "input operand" ); |
3649 | } |
3650 | |
3651 | XVECEXP (body, 0, i++) = gen_rtx_CLOBBER (VOIDmode, clobbered_reg); |
3652 | } |
3653 | |
3654 | if (nlabels > 0) |
3655 | emit_jump_insn (body); |
3656 | else |
3657 | emit_insn (body); |
3658 | } |
3659 | |
3660 | generating_concat_p = old_generating_concat_p; |
3661 | |
3662 | if (fallthru_label) |
3663 | emit_label (fallthru_label); |
3664 | |
3665 | if (after_md_seq) |
3666 | emit_insn (after_md_seq); |
3667 | if (after_rtl_seq) |
3668 | { |
3669 | if (nlabels == 0) |
3670 | emit_insn (after_rtl_seq); |
3671 | else |
3672 | { |
3673 | edge e; |
3674 | edge_iterator ei; |
3675 | unsigned int cnt = EDGE_COUNT (gimple_bb (stmt)->succs); |
3676 | |
3677 | FOR_EACH_EDGE (e, ei, gimple_bb (stmt)->succs) |
3678 | { |
3679 | rtx_insn *copy; |
3680 | if (--cnt == 0) |
3681 | copy = after_rtl_seq; |
3682 | else |
3683 | { |
3684 | start_sequence (); |
3685 | duplicate_insn_chain (after_rtl_seq, after_rtl_end, |
3686 | NULL, NULL); |
3687 | copy = get_insns (); |
3688 | end_sequence (); |
3689 | } |
3690 | prepend_insn_to_edge (copy, e); |
3691 | } |
3692 | } |
3693 | } |
3694 | |
3695 | free_temp_slots (); |
3696 | crtl->has_asm_statement = 1; |
3697 | } |
3698 | |
3699 | /* Emit code to jump to the address |
3700 | specified by the pointer expression EXP. */ |
3701 | |
3702 | static void |
3703 | expand_computed_goto (tree exp) |
3704 | { |
3705 | rtx x = expand_normal (exp); |
3706 | |
3707 | do_pending_stack_adjust (); |
3708 | emit_indirect_jump (x); |
3709 | } |
3710 | |
3711 | /* Generate RTL code for a `goto' statement with target label LABEL. |
3712 | LABEL should be a LABEL_DECL tree node that was or will later be |
3713 | defined with `expand_label'. */ |
3714 | |
3715 | static void |
3716 | expand_goto (tree label) |
3717 | { |
3718 | if (flag_checking) |
3719 | { |
3720 | /* Check for a nonlocal goto to a containing function. Should have |
3721 | gotten translated to __builtin_nonlocal_goto. */ |
3722 | tree context = decl_function_context (label); |
3723 | gcc_assert (!context || context == current_function_decl); |
3724 | } |
3725 | |
3726 | emit_jump (jump_target_rtx (label)); |
3727 | } |
3728 | |
3729 | /* Output a return with no value. */ |
3730 | |
3731 | static void |
3732 | expand_null_return_1 (void) |
3733 | { |
3734 | clear_pending_stack_adjust (); |
3735 | do_pending_stack_adjust (); |
3736 | emit_jump (return_label); |
3737 | } |
3738 | |
3739 | /* Generate RTL to return from the current function, with no value. |
3740 | (That is, we do not do anything about returning any value.) */ |
3741 | |
3742 | void |
3743 | expand_null_return (void) |
3744 | { |
3745 | /* If this function was declared to return a value, but we |
3746 | didn't, clobber the return registers so that they are not |
3747 | propagated live to the rest of the function. */ |
3748 | clobber_return_register (); |
3749 | |
3750 | expand_null_return_1 (); |
3751 | } |
3752 | |
3753 | /* Generate RTL to return from the current function, with value VAL. */ |
3754 | |
3755 | static void |
3756 | expand_value_return (rtx val) |
3757 | { |
3758 | /* Copy the value to the return location unless it's already there. */ |
3759 | |
3760 | tree decl = DECL_RESULT (current_function_decl); |
3761 | rtx return_reg = DECL_RTL (decl); |
3762 | if (return_reg != val) |
3763 | { |
3764 | tree funtype = TREE_TYPE (current_function_decl); |
3765 | tree type = TREE_TYPE (decl); |
3766 | int unsignedp = TYPE_UNSIGNED (type); |
3767 | machine_mode old_mode = DECL_MODE (decl); |
3768 | machine_mode mode; |
3769 | if (DECL_BY_REFERENCE (decl)) |
3770 | mode = promote_function_mode (type, old_mode, &unsignedp, funtype, 2); |
3771 | else |
3772 | mode = promote_function_mode (type, old_mode, &unsignedp, funtype, 1); |
3773 | |
3774 | if (mode != old_mode) |
3775 | { |
3776 | /* Some ABIs require scalar floating point modes to be returned |
3777 | in a wider scalar integer mode. We need to explicitly |
3778 | reinterpret to an integer mode of the correct precision |
3779 | before extending to the desired result. */ |
3780 | if (SCALAR_INT_MODE_P (mode) |
3781 | && SCALAR_FLOAT_MODE_P (old_mode) |
3782 | && known_gt (GET_MODE_SIZE (mode), GET_MODE_SIZE (old_mode))) |
3783 | val = convert_float_to_wider_int (mode, fmode: old_mode, x: val); |
3784 | else |
3785 | val = convert_modes (mode, oldmode: old_mode, x: val, unsignedp); |
3786 | } |
3787 | |
3788 | if (GET_CODE (return_reg) == PARALLEL) |
3789 | emit_group_load (return_reg, val, type, int_size_in_bytes (type)); |
3790 | else |
3791 | emit_move_insn (return_reg, val); |
3792 | } |
3793 | |
3794 | expand_null_return_1 (); |
3795 | } |
3796 | |
3797 | /* Generate RTL to evaluate the expression RETVAL and return it |
3798 | from the current function. */ |
3799 | |
3800 | static void |
3801 | expand_return (tree retval) |
3802 | { |
3803 | rtx result_rtl; |
3804 | rtx val = 0; |
3805 | tree retval_rhs; |
3806 | |
3807 | /* If function wants no value, give it none. */ |
3808 | if (VOID_TYPE_P (TREE_TYPE (TREE_TYPE (current_function_decl)))) |
3809 | { |
3810 | expand_normal (exp: retval); |
3811 | expand_null_return (); |
3812 | return; |
3813 | } |
3814 | |
3815 | if (retval == error_mark_node) |
3816 | { |
3817 | /* Treat this like a return of no value from a function that |
3818 | returns a value. */ |
3819 | expand_null_return (); |
3820 | return; |
3821 | } |
3822 | else if ((TREE_CODE (retval) == MODIFY_EXPR |
3823 | || TREE_CODE (retval) == INIT_EXPR) |
3824 | && TREE_CODE (TREE_OPERAND (retval, 0)) == RESULT_DECL) |
3825 | retval_rhs = TREE_OPERAND (retval, 1); |
3826 | else |
3827 | retval_rhs = retval; |
3828 | |
3829 | result_rtl = DECL_RTL (DECL_RESULT (current_function_decl)); |
3830 | |
3831 | /* If we are returning the RESULT_DECL, then the value has already |
3832 | been stored into it, so we don't have to do anything special. */ |
3833 | if (TREE_CODE (retval_rhs) == RESULT_DECL) |
3834 | expand_value_return (val: result_rtl); |
3835 | |
3836 | /* If the result is an aggregate that is being returned in one (or more) |
3837 | registers, load the registers here. */ |
3838 | |
3839 | else if (retval_rhs != 0 |
3840 | && TYPE_MODE (TREE_TYPE (retval_rhs)) == BLKmode |
3841 | && REG_P (result_rtl)) |
3842 | { |
3843 | val = copy_blkmode_to_reg (GET_MODE (result_rtl), retval_rhs); |
3844 | if (val) |
3845 | { |
3846 | /* Use the mode of the result value on the return register. */ |
3847 | PUT_MODE (x: result_rtl, GET_MODE (val)); |
3848 | expand_value_return (val); |
3849 | } |
3850 | else |
3851 | expand_null_return (); |
3852 | } |
3853 | else if (retval_rhs != 0 |
3854 | && !VOID_TYPE_P (TREE_TYPE (retval_rhs)) |
3855 | && (REG_P (result_rtl) |
3856 | || (GET_CODE (result_rtl) == PARALLEL))) |
3857 | { |
3858 | /* Compute the return value into a temporary (usually a pseudo reg). */ |
3859 | val |
3860 | = assign_temp (TREE_TYPE (DECL_RESULT (current_function_decl)), 0, 1); |
3861 | val = expand_expr (exp: retval_rhs, target: val, GET_MODE (val), modifier: EXPAND_NORMAL); |
3862 | val = force_not_mem (val); |
3863 | expand_value_return (val); |
3864 | } |
3865 | else |
3866 | { |
3867 | /* No hard reg used; calculate value into hard return reg. */ |
3868 | expand_expr (exp: retval, const0_rtx, VOIDmode, modifier: EXPAND_NORMAL); |
3869 | expand_value_return (val: result_rtl); |
3870 | } |
3871 | } |
3872 | |
3873 | /* Expand a clobber of LHS. If LHS is stored it in a multi-part |
3874 | register, tell the rtl optimizers that its value is no longer |
3875 | needed. */ |
3876 | |
3877 | static void |
3878 | expand_clobber (tree lhs) |
3879 | { |
3880 | if (DECL_P (lhs)) |
3881 | { |
3882 | rtx decl_rtl = DECL_RTL_IF_SET (lhs); |
3883 | if (decl_rtl && REG_P (decl_rtl)) |
3884 | { |
3885 | machine_mode decl_mode = GET_MODE (decl_rtl); |
3886 | if (maybe_gt (GET_MODE_SIZE (decl_mode), |
3887 | REGMODE_NATURAL_SIZE (decl_mode))) |
3888 | emit_clobber (decl_rtl); |
3889 | } |
3890 | } |
3891 | } |
3892 | |
3893 | /* A subroutine of expand_gimple_stmt, expanding one gimple statement |
3894 | STMT that doesn't require special handling for outgoing edges. That |
3895 | is no tailcalls and no GIMPLE_COND. */ |
3896 | |
3897 | static void |
3898 | expand_gimple_stmt_1 (gimple *stmt) |
3899 | { |
3900 | tree op0; |
3901 | |
3902 | set_curr_insn_location (gimple_location (g: stmt)); |
3903 | |
3904 | switch (gimple_code (g: stmt)) |
3905 | { |
3906 | case GIMPLE_GOTO: |
3907 | op0 = gimple_goto_dest (gs: stmt); |
3908 | if (TREE_CODE (op0) == LABEL_DECL) |
3909 | expand_goto (label: op0); |
3910 | else |
3911 | expand_computed_goto (exp: op0); |
3912 | break; |
3913 | case GIMPLE_LABEL: |
3914 | expand_label (gimple_label_label (gs: as_a <glabel *> (p: stmt))); |
3915 | break; |
3916 | case GIMPLE_NOP: |
3917 | case GIMPLE_PREDICT: |
3918 | break; |
3919 | case GIMPLE_SWITCH: |
3920 | { |
3921 | gswitch *swtch = as_a <gswitch *> (p: stmt); |
3922 | if (gimple_switch_num_labels (gs: swtch) == 1) |
3923 | expand_goto (CASE_LABEL (gimple_switch_default_label (swtch))); |
3924 | else |
3925 | expand_case (swtch); |
3926 | } |
3927 | break; |
3928 | case GIMPLE_ASM: |
3929 | expand_asm_stmt (stmt: as_a <gasm *> (p: stmt)); |
3930 | break; |
3931 | case GIMPLE_CALL: |
3932 | expand_call_stmt (stmt: as_a <gcall *> (p: stmt)); |
3933 | break; |
3934 | |
3935 | case GIMPLE_RETURN: |
3936 | { |
3937 | op0 = gimple_return_retval (gs: as_a <greturn *> (p: stmt)); |
3938 | |
3939 | /* If a return doesn't have a location, it very likely represents |
3940 | multiple user returns so we cannot let it inherit the location |
3941 | of the last statement of the previous basic block in RTL. */ |
3942 | if (!gimple_has_location (g: stmt)) |
3943 | set_curr_insn_location (cfun->function_end_locus); |
3944 | |
3945 | if (op0 && op0 != error_mark_node) |
3946 | { |
3947 | tree result = DECL_RESULT (current_function_decl); |
3948 | |
3949 | /* If we are not returning the current function's RESULT_DECL, |
3950 | build an assignment to it. */ |
3951 | if (op0 != result) |
3952 | { |
3953 | /* I believe that a function's RESULT_DECL is unique. */ |
3954 | gcc_assert (TREE_CODE (op0) != RESULT_DECL); |
3955 | |
3956 | /* ??? We'd like to use simply expand_assignment here, |
3957 | but this fails if the value is of BLKmode but the return |
3958 | decl is a register. expand_return has special handling |
3959 | for this combination, which eventually should move |
3960 | to common code. See comments there. Until then, let's |
3961 | build a modify expression :-/ */ |
3962 | op0 = build2 (MODIFY_EXPR, TREE_TYPE (result), |
3963 | result, op0); |
3964 | } |
3965 | } |
3966 | |
3967 | if (!op0) |
3968 | expand_null_return (); |
3969 | else |
3970 | expand_return (retval: op0); |
3971 | } |
3972 | break; |
3973 | |
3974 | case GIMPLE_ASSIGN: |
3975 | { |
3976 | gassign *assign_stmt = as_a <gassign *> (p: stmt); |
3977 | tree lhs = gimple_assign_lhs (gs: assign_stmt); |
3978 | |
3979 | /* Tree expand used to fiddle with |= and &= of two bitfield |
3980 | COMPONENT_REFs here. This can't happen with gimple, the LHS |
3981 | of binary assigns must be a gimple reg. */ |
3982 | |
3983 | if (TREE_CODE (lhs) != SSA_NAME |
3984 | || gimple_assign_rhs_class (gs: assign_stmt) == GIMPLE_SINGLE_RHS) |
3985 | { |
3986 | tree rhs = gimple_assign_rhs1 (gs: assign_stmt); |
3987 | gcc_assert (gimple_assign_rhs_class (assign_stmt) |
3988 | == GIMPLE_SINGLE_RHS); |
3989 | if (gimple_has_location (g: stmt) && CAN_HAVE_LOCATION_P (rhs) |
3990 | /* Do not put locations on possibly shared trees. */ |
3991 | && !is_gimple_min_invariant (rhs)) |
3992 | SET_EXPR_LOCATION (rhs, gimple_location (stmt)); |
3993 | if (TREE_CLOBBER_P (rhs)) |
3994 | /* This is a clobber to mark the going out of scope for |
3995 | this LHS. */ |
3996 | expand_clobber (lhs); |
3997 | else |
3998 | expand_assignment (lhs, rhs, |
3999 | gimple_assign_nontemporal_move_p ( |
4000 | gs: assign_stmt)); |
4001 | } |
4002 | else |
4003 | { |
4004 | rtx target, temp; |
4005 | bool nontemporal = gimple_assign_nontemporal_move_p (gs: assign_stmt); |
4006 | bool promoted = false; |
4007 | |
4008 | target = expand_expr (exp: lhs, NULL_RTX, VOIDmode, modifier: EXPAND_WRITE); |
4009 | if (GET_CODE (target) == SUBREG && SUBREG_PROMOTED_VAR_P (target)) |
4010 | promoted = true; |
4011 | |
4012 | /* If we want to use a nontemporal store, force the value to |
4013 | register first. If we store into a promoted register, |
4014 | don't directly expand to target. */ |
4015 | temp = nontemporal || promoted ? NULL_RTX : target; |
4016 | temp = expand_expr_real_gassign (assign_stmt, temp, |
4017 | GET_MODE (target), modifier: EXPAND_NORMAL); |
4018 | |
4019 | if (temp == target) |
4020 | ; |
4021 | else if (promoted) |
4022 | { |
4023 | int unsignedp = SUBREG_PROMOTED_SIGN (target); |
4024 | /* If TEMP is a VOIDmode constant, use convert_modes to make |
4025 | sure that we properly convert it. */ |
4026 | if (CONSTANT_P (temp) && GET_MODE (temp) == VOIDmode) |
4027 | { |
4028 | temp = convert_modes (GET_MODE (target), |
4029 | TYPE_MODE (TREE_TYPE (lhs)), |
4030 | x: temp, unsignedp); |
4031 | temp = convert_modes (GET_MODE (SUBREG_REG (target)), |
4032 | GET_MODE (target), x: temp, unsignedp); |
4033 | } |
4034 | |
4035 | convert_move (SUBREG_REG (target), temp, unsignedp); |
4036 | } |
4037 | else if (nontemporal && emit_storent_insn (to: target, from: temp)) |
4038 | ; |
4039 | else |
4040 | { |
4041 | temp = force_operand (temp, target); |
4042 | if (temp != target) |
4043 | emit_move_insn (target, temp); |
4044 | } |
4045 | } |
4046 | } |
4047 | break; |
4048 | |
4049 | default: |
4050 | gcc_unreachable (); |
4051 | } |
4052 | } |
4053 | |
4054 | /* Expand one gimple statement STMT and return the last RTL instruction |
4055 | before any of the newly generated ones. |
4056 | |
4057 | In addition to generating the necessary RTL instructions this also |
4058 | sets REG_EH_REGION notes if necessary and sets the current source |
4059 | location for diagnostics. */ |
4060 | |
4061 | static rtx_insn * |
4062 | expand_gimple_stmt (gimple *stmt) |
4063 | { |
4064 | location_t saved_location = input_location; |
4065 | rtx_insn *last = get_last_insn (); |
4066 | int lp_nr; |
4067 | |
4068 | gcc_assert (cfun); |
4069 | |
4070 | /* We need to save and restore the current source location so that errors |
4071 | discovered during expansion are emitted with the right location. But |
4072 | it would be better if the diagnostic routines used the source location |
4073 | embedded in the tree nodes rather than globals. */ |
4074 | if (gimple_has_location (g: stmt)) |
4075 | input_location = gimple_location (g: stmt); |
4076 | |
4077 | expand_gimple_stmt_1 (stmt); |
4078 | |
4079 | /* Free any temporaries used to evaluate this statement. */ |
4080 | free_temp_slots (); |
4081 | |
4082 | input_location = saved_location; |
4083 | |
4084 | /* Mark all insns that may trap. */ |
4085 | lp_nr = lookup_stmt_eh_lp (stmt); |
4086 | if (lp_nr) |
4087 | { |
4088 | rtx_insn *insn; |
4089 | for (insn = next_real_insn (last); insn; |
4090 | insn = next_real_insn (insn)) |
4091 | { |
4092 | if (! find_reg_note (insn, REG_EH_REGION, NULL_RTX) |
4093 | /* If we want exceptions for non-call insns, any |
4094 | may_trap_p instruction may throw. */ |
4095 | && GET_CODE (PATTERN (insn)) != CLOBBER |
4096 | && GET_CODE (PATTERN (insn)) != USE |
4097 | && insn_could_throw_p (insn)) |
4098 | make_reg_eh_region_note (insn, ecf_flags: 0, lp_nr); |
4099 | } |
4100 | } |
4101 | |
4102 | return last; |
4103 | } |
4104 | |
4105 | /* A subroutine of expand_gimple_basic_block. Expand one GIMPLE_CALL |
4106 | that has CALL_EXPR_TAILCALL set. Returns non-null if we actually |
4107 | generated a tail call (something that might be denied by the ABI |
4108 | rules governing the call; see calls.cc). |
4109 | |
4110 | Sets CAN_FALLTHRU if we generated a *conditional* tail call, and |
4111 | can still reach the rest of BB. The case here is __builtin_sqrt, |
4112 | where the NaN result goes through the external function (with a |
4113 | tailcall) and the normal result happens via a sqrt instruction. */ |
4114 | |
4115 | static basic_block |
4116 | expand_gimple_tailcall (basic_block bb, gcall *stmt, bool *can_fallthru) |
4117 | { |
4118 | rtx_insn *last2, *last; |
4119 | edge e; |
4120 | edge_iterator ei; |
4121 | profile_probability probability; |
4122 | |
4123 | last2 = last = expand_gimple_stmt (stmt); |
4124 | |
4125 | for (last = NEXT_INSN (insn: last); last; last = NEXT_INSN (insn: last)) |
4126 | if (CALL_P (last) && SIBLING_CALL_P (last)) |
4127 | goto found; |
4128 | |
4129 | maybe_dump_rtl_for_gimple_stmt (stmt, since: last2); |
4130 | |
4131 | *can_fallthru = true; |
4132 | return NULL; |
4133 | |
4134 | found: |
4135 | /* ??? Wouldn't it be better to just reset any pending stack adjust? |
4136 | Any instructions emitted here are about to be deleted. */ |
4137 | do_pending_stack_adjust (); |
4138 | |
4139 | /* Remove any non-eh, non-abnormal edges that don't go to exit. */ |
4140 | /* ??? I.e. the fallthrough edge. HOWEVER! If there were to be |
4141 | EH or abnormal edges, we shouldn't have created a tail call in |
4142 | the first place. So it seems to me we should just be removing |
4143 | all edges here, or redirecting the existing fallthru edge to |
4144 | the exit block. */ |
4145 | |
4146 | probability = profile_probability::never (); |
4147 | |
4148 | for (ei = ei_start (bb->succs); (e = ei_safe_edge (i: ei)); ) |
4149 | { |
4150 | if (!(e->flags & (EDGE_ABNORMAL | EDGE_EH))) |
4151 | { |
4152 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
4153 | e->dest->count -= e->count (); |
4154 | probability += e->probability; |
4155 | remove_edge (e); |
4156 | } |
4157 | else |
4158 | ei_next (i: &ei); |
4159 | } |
4160 | |
4161 | /* This is somewhat ugly: the call_expr expander often emits instructions |
4162 | after the sibcall (to perform the function return). These confuse the |
4163 | find_many_sub_basic_blocks code, so we need to get rid of these. */ |
4164 | last = NEXT_INSN (insn: last); |
4165 | gcc_assert (BARRIER_P (last)); |
4166 | |
4167 | *can_fallthru = false; |
4168 | while (NEXT_INSN (insn: last)) |
4169 | { |
4170 | /* For instance an sqrt builtin expander expands if with |
4171 | sibcall in the then and label for `else`. */ |
4172 | if (LABEL_P (NEXT_INSN (last))) |
4173 | { |
4174 | *can_fallthru = true; |
4175 | break; |
4176 | } |
4177 | delete_insn (NEXT_INSN (insn: last)); |
4178 | } |
4179 | |
4180 | e = make_edge (bb, EXIT_BLOCK_PTR_FOR_FN (cfun), EDGE_ABNORMAL |
4181 | | EDGE_SIBCALL); |
4182 | e->probability = probability; |
4183 | BB_END (bb) = last; |
4184 | update_bb_for_insn (bb); |
4185 | |
4186 | if (NEXT_INSN (insn: last)) |
4187 | { |
4188 | bb = create_basic_block (NEXT_INSN (insn: last), get_last_insn (), bb); |
4189 | |
4190 | last = BB_END (bb); |
4191 | if (BARRIER_P (last)) |
4192 | BB_END (bb) = PREV_INSN (insn: last); |
4193 | } |
4194 | |
4195 | maybe_dump_rtl_for_gimple_stmt (stmt, since: last2); |
4196 | |
4197 | return bb; |
4198 | } |
4199 | |
4200 | /* Return the difference between the floor and the truncated result of |
4201 | a signed division by OP1 with remainder MOD. */ |
4202 | static rtx |
4203 | floor_sdiv_adjust (machine_mode mode, rtx mod, rtx op1) |
4204 | { |
4205 | /* (mod != 0 ? (op1 / mod < 0 ? -1 : 0) : 0) */ |
4206 | return gen_rtx_IF_THEN_ELSE |
4207 | (mode, gen_rtx_NE (BImode, mod, const0_rtx), |
4208 | gen_rtx_IF_THEN_ELSE |
4209 | (mode, gen_rtx_LT (BImode, |
4210 | gen_rtx_DIV (mode, op1, mod), |
4211 | const0_rtx), |
4212 | constm1_rtx, const0_rtx), |
4213 | const0_rtx); |
4214 | } |
4215 | |
4216 | /* Return the difference between the ceil and the truncated result of |
4217 | a signed division by OP1 with remainder MOD. */ |
4218 | static rtx |
4219 | ceil_sdiv_adjust (machine_mode mode, rtx mod, rtx op1) |
4220 | { |
4221 | /* (mod != 0 ? (op1 / mod > 0 ? 1 : 0) : 0) */ |
4222 | return gen_rtx_IF_THEN_ELSE |
4223 | (mode, gen_rtx_NE (BImode, mod, const0_rtx), |
4224 | gen_rtx_IF_THEN_ELSE |
4225 | (mode, gen_rtx_GT (BImode, |
4226 | gen_rtx_DIV (mode, op1, mod), |
4227 | const0_rtx), |
4228 | const1_rtx, const0_rtx), |
4229 | const0_rtx); |
4230 | } |
4231 | |
4232 | /* Return the difference between the ceil and the truncated result of |
4233 | an unsigned division by OP1 with remainder MOD. */ |
4234 | static rtx |
4235 | ceil_udiv_adjust (machine_mode mode, rtx mod, rtx op1 ATTRIBUTE_UNUSED) |
4236 | { |
4237 | /* (mod != 0 ? 1 : 0) */ |
4238 | return gen_rtx_IF_THEN_ELSE |
4239 | (mode, gen_rtx_NE (BImode, mod, const0_rtx), |
4240 | const1_rtx, const0_rtx); |
4241 | } |
4242 | |
4243 | /* Return the difference between the rounded and the truncated result |
4244 | of a signed division by OP1 with remainder MOD. Halfway cases are |
4245 | rounded away from zero, rather than to the nearest even number. */ |
4246 | static rtx |
4247 | round_sdiv_adjust (machine_mode mode, rtx mod, rtx op1) |
4248 | { |
4249 | /* (abs (mod) >= abs (op1) - abs (mod) |
4250 | ? (op1 / mod > 0 ? 1 : -1) |
4251 | : 0) */ |
4252 | return gen_rtx_IF_THEN_ELSE |
4253 | (mode, gen_rtx_GE (BImode, gen_rtx_ABS (mode, mod), |
4254 | gen_rtx_MINUS (mode, |
4255 | gen_rtx_ABS (mode, op1), |
4256 | gen_rtx_ABS (mode, mod))), |
4257 | gen_rtx_IF_THEN_ELSE |
4258 | (mode, gen_rtx_GT (BImode, |
4259 | gen_rtx_DIV (mode, op1, mod), |
4260 | const0_rtx), |
4261 | const1_rtx, constm1_rtx), |
4262 | const0_rtx); |
4263 | } |
4264 | |
4265 | /* Return the difference between the rounded and the truncated result |
4266 | of a unsigned division by OP1 with remainder MOD. Halfway cases |
4267 | are rounded away from zero, rather than to the nearest even |
4268 | number. */ |
4269 | static rtx |
4270 | round_udiv_adjust (machine_mode mode, rtx mod, rtx op1) |
4271 | { |
4272 | /* (mod >= op1 - mod ? 1 : 0) */ |
4273 | return gen_rtx_IF_THEN_ELSE |
4274 | (mode, gen_rtx_GE (BImode, mod, |
4275 | gen_rtx_MINUS (mode, op1, mod)), |
4276 | const1_rtx, const0_rtx); |
4277 | } |
4278 | |
4279 | /* Convert X to MODE, that must be Pmode or ptr_mode, without emitting |
4280 | any rtl. */ |
4281 | |
4282 | static rtx |
4283 | convert_debug_memory_address (scalar_int_mode mode, rtx x, |
4284 | addr_space_t as) |
4285 | { |
4286 | #ifndef POINTERS_EXTEND_UNSIGNED |
4287 | gcc_assert (mode == Pmode |
4288 | || mode == targetm.addr_space.address_mode (as)); |
4289 | gcc_assert (GET_MODE (x) == mode || GET_MODE (x) == VOIDmode); |
4290 | #else |
4291 | rtx temp; |
4292 | |
4293 | gcc_assert (targetm.addr_space.valid_pointer_mode (mode, as)); |
4294 | |
4295 | if (GET_MODE (x) == mode || GET_MODE (x) == VOIDmode) |
4296 | return x; |
4297 | |
4298 | /* X must have some form of address mode already. */ |
4299 | scalar_int_mode xmode = as_a <scalar_int_mode> (GET_MODE (x)); |
4300 | if (GET_MODE_PRECISION (mode) < GET_MODE_PRECISION (mode: xmode)) |
4301 | x = lowpart_subreg (outermode: mode, op: x, innermode: xmode); |
4302 | else if (POINTERS_EXTEND_UNSIGNED > 0) |
4303 | x = gen_rtx_ZERO_EXTEND (mode, x); |
4304 | else if (!POINTERS_EXTEND_UNSIGNED) |
4305 | x = gen_rtx_SIGN_EXTEND (mode, x); |
4306 | else |
4307 | { |
4308 | switch (GET_CODE (x)) |
4309 | { |
4310 | case SUBREG: |
4311 | if ((SUBREG_PROMOTED_VAR_P (x) |
4312 | || (REG_P (SUBREG_REG (x)) && REG_POINTER (SUBREG_REG (x))) |
4313 | || (GET_CODE (SUBREG_REG (x)) == PLUS |
4314 | && REG_P (XEXP (SUBREG_REG (x), 0)) |
4315 | && REG_POINTER (XEXP (SUBREG_REG (x), 0)) |
4316 | && CONST_INT_P (XEXP (SUBREG_REG (x), 1)))) |
4317 | && GET_MODE (SUBREG_REG (x)) == mode) |
4318 | return SUBREG_REG (x); |
4319 | break; |
4320 | case LABEL_REF: |
4321 | temp = gen_rtx_LABEL_REF (mode, label_ref_label (x)); |
4322 | LABEL_REF_NONLOCAL_P (temp) = LABEL_REF_NONLOCAL_P (x); |
4323 | return temp; |
4324 | case SYMBOL_REF: |
4325 | temp = shallow_copy_rtx (x); |
4326 | PUT_MODE (x: temp, mode); |
4327 | return temp; |
4328 | case CONST: |
4329 | temp = convert_debug_memory_address (mode, XEXP (x, 0), as); |
4330 | if (temp) |
4331 | temp = gen_rtx_CONST (mode, temp); |
4332 | return temp; |
4333 | case PLUS: |
4334 | case MINUS: |
4335 | if (CONST_INT_P (XEXP (x, 1))) |
4336 | { |
4337 | temp = convert_debug_memory_address (mode, XEXP (x, 0), as); |
4338 | if (temp) |
4339 | return gen_rtx_fmt_ee (GET_CODE (x), mode, temp, XEXP (x, 1)); |
4340 | } |
4341 | break; |
4342 | default: |
4343 | break; |
4344 | } |
4345 | /* Don't know how to express ptr_extend as operation in debug info. */ |
4346 | return NULL; |
4347 | } |
4348 | #endif /* POINTERS_EXTEND_UNSIGNED */ |
4349 | |
4350 | return x; |
4351 | } |
4352 | |
4353 | /* Map from SSA_NAMEs to corresponding DEBUG_EXPR_DECLs created |
4354 | by avoid_deep_ter_for_debug. */ |
4355 | |
4356 | static hash_map<tree, tree> *deep_ter_debug_map; |
4357 | |
4358 | /* Split too deep TER chains for debug stmts using debug temporaries. */ |
4359 | |
4360 | static void |
4361 | avoid_deep_ter_for_debug (gimple *stmt, int depth) |
4362 | { |
4363 | use_operand_p use_p; |
4364 | ssa_op_iter iter; |
4365 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) |
4366 | { |
4367 | tree use = USE_FROM_PTR (use_p); |
4368 | if (TREE_CODE (use) != SSA_NAME || SSA_NAME_IS_DEFAULT_DEF (use)) |
4369 | continue; |
4370 | gimple *g = get_gimple_for_ssa_name (exp: use); |
4371 | if (g == NULL) |
4372 | continue; |
4373 | if (depth > 6 && !stmt_ends_bb_p (g)) |
4374 | { |
4375 | if (deep_ter_debug_map == NULL) |
4376 | deep_ter_debug_map = new hash_map<tree, tree>; |
4377 | |
4378 | tree &vexpr = deep_ter_debug_map->get_or_insert (k: use); |
4379 | if (vexpr != NULL) |
4380 | continue; |
4381 | vexpr = build_debug_expr_decl (TREE_TYPE (use)); |
4382 | gimple *def_temp = gimple_build_debug_bind (vexpr, use, g); |
4383 | gimple_stmt_iterator gsi = gsi_for_stmt (g); |
4384 | gsi_insert_after (&gsi, def_temp, GSI_NEW_STMT); |
4385 | avoid_deep_ter_for_debug (stmt: def_temp, depth: 0); |
4386 | } |
4387 | else |
4388 | avoid_deep_ter_for_debug (stmt: g, depth: depth + 1); |
4389 | } |
4390 | } |
4391 | |
4392 | /* Return an RTX equivalent to the value of the parameter DECL. */ |
4393 | |
4394 | static rtx |
4395 | expand_debug_parm_decl (tree decl) |
4396 | { |
4397 | rtx incoming = DECL_INCOMING_RTL (decl); |
4398 | |
4399 | if (incoming |
4400 | && GET_MODE (incoming) != BLKmode |
4401 | && ((REG_P (incoming) && HARD_REGISTER_P (incoming)) |
4402 | || (MEM_P (incoming) |
4403 | && REG_P (XEXP (incoming, 0)) |
4404 | && HARD_REGISTER_P (XEXP (incoming, 0))))) |
4405 | { |
4406 | rtx rtl = gen_rtx_ENTRY_VALUE (GET_MODE (incoming)); |
4407 | |
4408 | #ifdef HAVE_window_save |
4409 | /* DECL_INCOMING_RTL uses the INCOMING_REGNO of parameter registers. |
4410 | If the target machine has an explicit window save instruction, the |
4411 | actual entry value is the corresponding OUTGOING_REGNO instead. */ |
4412 | if (REG_P (incoming) |
4413 | && OUTGOING_REGNO (REGNO (incoming)) != REGNO (incoming)) |
4414 | incoming |
4415 | = gen_rtx_REG_offset (incoming, GET_MODE (incoming), |
4416 | OUTGOING_REGNO (REGNO (incoming)), 0); |
4417 | else if (MEM_P (incoming)) |
4418 | { |
4419 | rtx reg = XEXP (incoming, 0); |
4420 | if (OUTGOING_REGNO (REGNO (reg)) != REGNO (reg)) |
4421 | { |
4422 | reg = gen_raw_REG (GET_MODE (reg), OUTGOING_REGNO (REGNO (reg))); |
4423 | incoming = replace_equiv_address_nv (incoming, reg); |
4424 | } |
4425 | else |
4426 | incoming = copy_rtx (incoming); |
4427 | } |
4428 | #endif |
4429 | |
4430 | ENTRY_VALUE_EXP (rtl) = incoming; |
4431 | return rtl; |
4432 | } |
4433 | |
4434 | if (incoming |
4435 | && GET_MODE (incoming) != BLKmode |
4436 | && !TREE_ADDRESSABLE (decl) |
4437 | && MEM_P (incoming) |
4438 | && (XEXP (incoming, 0) == virtual_incoming_args_rtx |
4439 | || (GET_CODE (XEXP (incoming, 0)) == PLUS |
4440 | && XEXP (XEXP (incoming, 0), 0) == virtual_incoming_args_rtx |
4441 | && CONST_INT_P (XEXP (XEXP (incoming, 0), 1))))) |
4442 | return copy_rtx (incoming); |
4443 | |
4444 | return NULL_RTX; |
4445 | } |
4446 | |
4447 | /* Return an RTX equivalent to the value of the tree expression EXP. */ |
4448 | |
4449 | static rtx |
4450 | expand_debug_expr (tree exp) |
4451 | { |
4452 | rtx op0 = NULL_RTX, op1 = NULL_RTX, op2 = NULL_RTX; |
4453 | machine_mode mode = TYPE_MODE (TREE_TYPE (exp)); |
4454 | machine_mode inner_mode = VOIDmode; |
4455 | int unsignedp = TYPE_UNSIGNED (TREE_TYPE (exp)); |
4456 | addr_space_t as; |
4457 | scalar_int_mode op0_mode, op1_mode, addr_mode; |
4458 | |
4459 | switch (TREE_CODE_CLASS (TREE_CODE (exp))) |
4460 | { |
4461 | case tcc_expression: |
4462 | switch (TREE_CODE (exp)) |
4463 | { |
4464 | case COND_EXPR: |
4465 | case DOT_PROD_EXPR: |
4466 | case SAD_EXPR: |
4467 | case WIDEN_MULT_PLUS_EXPR: |
4468 | case WIDEN_MULT_MINUS_EXPR: |
4469 | goto ternary; |
4470 | |
4471 | case TRUTH_ANDIF_EXPR: |
4472 | case TRUTH_ORIF_EXPR: |
4473 | case TRUTH_AND_EXPR: |
4474 | case TRUTH_OR_EXPR: |
4475 | case TRUTH_XOR_EXPR: |
4476 | goto binary; |
4477 | |
4478 | case TRUTH_NOT_EXPR: |
4479 | goto unary; |
4480 | |
4481 | default: |
4482 | break; |
4483 | } |
4484 | break; |
4485 | |
4486 | ternary: |
4487 | op2 = expand_debug_expr (TREE_OPERAND (exp, 2)); |
4488 | if (!op2) |
4489 | return NULL_RTX; |
4490 | /* Fall through. */ |
4491 | |
4492 | binary: |
4493 | case tcc_binary: |
4494 | if (mode == BLKmode) |
4495 | return NULL_RTX; |
4496 | op1 = expand_debug_expr (TREE_OPERAND (exp, 1)); |
4497 | if (!op1) |
4498 | return NULL_RTX; |
4499 | switch (TREE_CODE (exp)) |
4500 | { |
4501 | case LSHIFT_EXPR: |
4502 | case RSHIFT_EXPR: |
4503 | case LROTATE_EXPR: |
4504 | case RROTATE_EXPR: |
4505 | case WIDEN_LSHIFT_EXPR: |
4506 | /* Ensure second operand isn't wider than the first one. */ |
4507 | inner_mode = TYPE_MODE (TREE_TYPE (TREE_OPERAND (exp, 1))); |
4508 | if (is_a <scalar_int_mode> (m: inner_mode, result: &op1_mode) |
4509 | && (GET_MODE_UNIT_PRECISION (mode) |
4510 | < GET_MODE_PRECISION (mode: op1_mode))) |
4511 | op1 = lowpart_subreg (GET_MODE_INNER (mode), op: op1, innermode: op1_mode); |
4512 | break; |
4513 | default: |
4514 | break; |
4515 | } |
4516 | /* Fall through. */ |
4517 | |
4518 | unary: |
4519 | case tcc_unary: |
4520 | if (mode == BLKmode) |
4521 | return NULL_RTX; |
4522 | inner_mode = TYPE_MODE (TREE_TYPE (TREE_OPERAND (exp, 0))); |
4523 | op0 = expand_debug_expr (TREE_OPERAND (exp, 0)); |
4524 | if (!op0) |
4525 | return NULL_RTX; |
4526 | break; |
4527 | |
4528 | case tcc_comparison: |
4529 | unsignedp = TYPE_UNSIGNED (TREE_TYPE (TREE_OPERAND (exp, 0))); |
4530 | goto binary; |
4531 | |
4532 | case tcc_type: |
4533 | case tcc_statement: |
4534 | gcc_unreachable (); |
4535 | |
4536 | case tcc_constant: |
4537 | case tcc_exceptional: |
4538 | case tcc_declaration: |
4539 | case tcc_reference: |
4540 | case tcc_vl_exp: |
4541 | break; |
4542 | } |
4543 | |
4544 | switch (TREE_CODE (exp)) |
4545 | { |
4546 | case STRING_CST: |
4547 | if (!lookup_constant_def (exp)) |
4548 | { |
4549 | if (strlen (TREE_STRING_POINTER (exp)) + 1 |
4550 | != (size_t) TREE_STRING_LENGTH (exp)) |
4551 | return NULL_RTX; |
4552 | op0 = gen_rtx_CONST_STRING (Pmode, TREE_STRING_POINTER (exp)); |
4553 | op0 = gen_rtx_MEM (BLKmode, op0); |
4554 | set_mem_attributes (op0, exp, 0); |
4555 | return op0; |
4556 | } |
4557 | /* Fall through. */ |
4558 | |
4559 | case INTEGER_CST: |
4560 | if (TREE_CODE (TREE_TYPE (exp)) == BITINT_TYPE |
4561 | && TYPE_MODE (TREE_TYPE (exp)) == BLKmode) |
4562 | return NULL; |
4563 | /* FALLTHRU */ |
4564 | case REAL_CST: |
4565 | case FIXED_CST: |
4566 | op0 = expand_expr (exp, NULL_RTX, mode, modifier: EXPAND_INITIALIZER); |
4567 | return op0; |
4568 | |
4569 | case POLY_INT_CST: |
4570 | return immed_wide_int_const (poly_int_cst_value (x: exp), mode); |
4571 | |
4572 | case COMPLEX_CST: |
4573 | gcc_assert (COMPLEX_MODE_P (mode)); |
4574 | op0 = expand_debug_expr (TREE_REALPART (exp)); |
4575 | op1 = expand_debug_expr (TREE_IMAGPART (exp)); |
4576 | return gen_rtx_CONCAT (mode, op0, op1); |
4577 | |
4578 | case DEBUG_EXPR_DECL: |
4579 | op0 = DECL_RTL_IF_SET (exp); |
4580 | |
4581 | if (op0) |
4582 | { |
4583 | if (GET_MODE (op0) != mode) |
4584 | gcc_assert (VECTOR_TYPE_P (TREE_TYPE (exp))); |
4585 | else |
4586 | return op0; |
4587 | } |
4588 | |
4589 | op0 = gen_rtx_DEBUG_EXPR (mode); |
4590 | DEBUG_EXPR_TREE_DECL (op0) = exp; |
4591 | SET_DECL_RTL (exp, op0); |
4592 | |
4593 | return op0; |
4594 | |
4595 | case VAR_DECL: |
4596 | case PARM_DECL: |
4597 | case FUNCTION_DECL: |
4598 | case LABEL_DECL: |
4599 | case CONST_DECL: |
4600 | case RESULT_DECL: |
4601 | op0 = DECL_RTL_IF_SET (exp); |
4602 | |
4603 | /* This decl was probably optimized away. */ |
4604 | if (!op0 |
4605 | /* At least label RTXen are sometimes replaced by |
4606 | NOTE_INSN_DELETED_LABEL. Any notes here are not |
4607 | handled by copy_rtx. */ |
4608 | || NOTE_P (op0)) |
4609 | { |
4610 | if (!VAR_P (exp) |
4611 | || DECL_EXTERNAL (exp) |
4612 | || !TREE_STATIC (exp) |
4613 | || !DECL_NAME (exp) |
4614 | || DECL_HARD_REGISTER (exp) |
4615 | || DECL_IN_CONSTANT_POOL (exp) |
4616 | || mode == VOIDmode |
4617 | || symtab_node::get (decl: exp) == NULL) |
4618 | return NULL; |
4619 | |
4620 | op0 = make_decl_rtl_for_debug (exp); |
4621 | if (!MEM_P (op0) |
4622 | || GET_CODE (XEXP (op0, 0)) != SYMBOL_REF |
4623 | || SYMBOL_REF_DECL (XEXP (op0, 0)) != exp) |
4624 | return NULL; |
4625 | } |
4626 | else if (VAR_P (exp) |
4627 | && is_global_var (t: exp) |
4628 | && symtab_node::get (decl: exp) == NULL) |
4629 | return NULL; |
4630 | else |
4631 | op0 = copy_rtx (op0); |
4632 | |
4633 | if (GET_MODE (op0) == BLKmode |
4634 | /* If op0 is not BLKmode, but mode is, adjust_mode |
4635 | below would ICE. While it is likely a FE bug, |
4636 | try to be robust here. See PR43166. */ |
4637 | || mode == BLKmode |
4638 | || (mode == VOIDmode && GET_MODE (op0) != VOIDmode)) |
4639 | { |
4640 | gcc_assert (MEM_P (op0)); |
4641 | op0 = adjust_address_nv (op0, mode, 0); |
4642 | return op0; |
4643 | } |
4644 | |
4645 | /* Fall through. */ |
4646 | |
4647 | adjust_mode: |
4648 | case PAREN_EXPR: |
4649 | CASE_CONVERT: |
4650 | { |
4651 | inner_mode = GET_MODE (op0); |
4652 | |
4653 | if (mode == inner_mode) |
4654 | return op0; |
4655 | |
4656 | if (inner_mode == VOIDmode) |
4657 | { |
4658 | if (TREE_CODE (exp) == SSA_NAME) |
4659 | inner_mode = TYPE_MODE (TREE_TYPE (exp)); |
4660 | else |
4661 | inner_mode = TYPE_MODE (TREE_TYPE (TREE_OPERAND (exp, 0))); |
4662 | if (mode == inner_mode) |
4663 | return op0; |
4664 | } |
4665 | |
4666 | if (FLOAT_MODE_P (mode) && FLOAT_MODE_P (inner_mode)) |
4667 | { |
4668 | if (GET_MODE_UNIT_BITSIZE (mode) |
4669 | == GET_MODE_UNIT_BITSIZE (inner_mode)) |
4670 | op0 = simplify_gen_subreg (outermode: mode, op: op0, innermode: inner_mode, byte: 0); |
4671 | else if (GET_MODE_UNIT_BITSIZE (mode) |
4672 | < GET_MODE_UNIT_BITSIZE (inner_mode)) |
4673 | op0 = simplify_gen_unary (code: FLOAT_TRUNCATE, mode, op: op0, op_mode: inner_mode); |
4674 | else |
4675 | op0 = simplify_gen_unary (code: FLOAT_EXTEND, mode, op: op0, op_mode: inner_mode); |
4676 | } |
4677 | else if (FLOAT_MODE_P (mode)) |
4678 | { |
4679 | gcc_assert (TREE_CODE (exp) != SSA_NAME); |
4680 | if (TYPE_UNSIGNED (TREE_TYPE (TREE_OPERAND (exp, 0)))) |
4681 | op0 = simplify_gen_unary (code: UNSIGNED_FLOAT, mode, op: op0, op_mode: inner_mode); |
4682 | else |
4683 | op0 = simplify_gen_unary (code: FLOAT, mode, op: op0, op_mode: inner_mode); |
4684 | } |
4685 | else if (FLOAT_MODE_P (inner_mode)) |
4686 | { |
4687 | if (unsignedp) |
4688 | op0 = simplify_gen_unary (code: UNSIGNED_FIX, mode, op: op0, op_mode: inner_mode); |
4689 | else |
4690 | op0 = simplify_gen_unary (code: FIX, mode, op: op0, op_mode: inner_mode); |
4691 | } |
4692 | else if (GET_MODE_UNIT_PRECISION (mode) |
4693 | == GET_MODE_UNIT_PRECISION (inner_mode)) |
4694 | op0 = lowpart_subreg (outermode: mode, op: op0, innermode: inner_mode); |
4695 | else if (GET_MODE_UNIT_PRECISION (mode) |
4696 | < GET_MODE_UNIT_PRECISION (inner_mode)) |
4697 | op0 = simplify_gen_unary (code: TRUNCATE, mode, op: op0, op_mode: inner_mode); |
4698 | else if (UNARY_CLASS_P (exp) |
4699 | ? TYPE_UNSIGNED (TREE_TYPE (TREE_OPERAND (exp, 0))) |
4700 | : unsignedp) |
4701 | op0 = simplify_gen_unary (code: ZERO_EXTEND, mode, op: op0, op_mode: inner_mode); |
4702 | else |
4703 | op0 = simplify_gen_unary (code: SIGN_EXTEND, mode, op: op0, op_mode: inner_mode); |
4704 | |
4705 | return op0; |
4706 | } |
4707 | |
4708 | case MEM_REF: |
4709 | if (!is_gimple_mem_ref_addr (TREE_OPERAND (exp, 0))) |
4710 | { |
4711 | tree newexp = fold_binary (MEM_REF, TREE_TYPE (exp), |
4712 | TREE_OPERAND (exp, 0), |
4713 | TREE_OPERAND (exp, 1)); |
4714 | if (newexp) |
4715 | return expand_debug_expr (exp: newexp); |
4716 | } |
4717 | /* FALLTHROUGH */ |
4718 | case INDIRECT_REF: |
4719 | inner_mode = TYPE_MODE (TREE_TYPE (TREE_OPERAND (exp, 0))); |
4720 | op0 = expand_debug_expr (TREE_OPERAND (exp, 0)); |
4721 | if (!op0) |
4722 | return NULL; |
4723 | |
4724 | if (TREE_CODE (exp) == MEM_REF) |
4725 | { |
4726 | if (GET_CODE (op0) == DEBUG_IMPLICIT_PTR |
4727 | || (GET_CODE (op0) == PLUS |
4728 | && GET_CODE (XEXP (op0, 0)) == DEBUG_IMPLICIT_PTR)) |
4729 | /* (mem (debug_implicit_ptr)) might confuse aliasing. |
4730 | Instead just use get_inner_reference. */ |
4731 | goto component_ref; |
4732 | |
4733 | op1 = expand_debug_expr (TREE_OPERAND (exp, 1)); |
4734 | poly_int64 offset; |
4735 | if (!op1 || !poly_int_rtx_p (x: op1, res: &offset)) |
4736 | return NULL; |
4737 | |
4738 | op0 = plus_constant (inner_mode, op0, offset); |
4739 | } |
4740 | |
4741 | as = TYPE_ADDR_SPACE (TREE_TYPE (TREE_TYPE (TREE_OPERAND (exp, 0)))); |
4742 | |
4743 | op0 = convert_debug_memory_address (mode: targetm.addr_space.address_mode (as), |
4744 | x: op0, as); |
4745 | if (op0 == NULL_RTX) |
4746 | return NULL; |
4747 | |
4748 | op0 = gen_rtx_MEM (mode, op0); |
4749 | set_mem_attributes (op0, exp, 0); |
4750 | if (TREE_CODE (exp) == MEM_REF |
4751 | && !is_gimple_mem_ref_addr (TREE_OPERAND (exp, 0))) |
4752 | set_mem_expr (op0, NULL_TREE); |
4753 | set_mem_addr_space (op0, as); |
4754 | |
4755 | return op0; |
4756 | |
4757 | case TARGET_MEM_REF: |
4758 | if (TREE_CODE (TMR_BASE (exp)) == ADDR_EXPR |
4759 | && !DECL_RTL_SET_P (TREE_OPERAND (TMR_BASE (exp), 0))) |
4760 | return NULL; |
4761 | |
4762 | op0 = expand_debug_expr |
4763 | (exp: tree_mem_ref_addr (build_pointer_type (TREE_TYPE (exp)), exp)); |
4764 | if (!op0) |
4765 | return NULL; |
4766 | |
4767 | as = TYPE_ADDR_SPACE (TREE_TYPE (TREE_TYPE (TREE_OPERAND (exp, 0)))); |
4768 | op0 = convert_debug_memory_address (mode: targetm.addr_space.address_mode (as), |
4769 | x: op0, as); |
4770 | if (op0 == NULL_RTX) |
4771 | return NULL; |
4772 | |
4773 | op0 = gen_rtx_MEM (mode, op0); |
4774 | |
4775 | set_mem_attributes (op0, exp, 0); |
4776 | set_mem_addr_space (op0, as); |
4777 | |
4778 | return op0; |
4779 | |
4780 | component_ref: |
4781 | case ARRAY_REF: |
4782 | case ARRAY_RANGE_REF: |
4783 | case COMPONENT_REF: |
4784 | case BIT_FIELD_REF: |
4785 | case REALPART_EXPR: |
4786 | case IMAGPART_EXPR: |
4787 | case VIEW_CONVERT_EXPR: |
4788 | { |
4789 | machine_mode mode1; |
4790 | poly_int64 bitsize, bitpos; |
4791 | tree offset; |
4792 | int reversep, volatilep = 0; |
4793 | tree tem |
4794 | = get_inner_reference (exp, &bitsize, &bitpos, &offset, &mode1, |
4795 | &unsignedp, &reversep, &volatilep); |
4796 | rtx orig_op0; |
4797 | |
4798 | if (known_eq (bitsize, 0)) |
4799 | return NULL; |
4800 | |
4801 | orig_op0 = op0 = expand_debug_expr (exp: tem); |
4802 | |
4803 | if (!op0) |
4804 | return NULL; |
4805 | |
4806 | if (offset) |
4807 | { |
4808 | machine_mode addrmode, offmode; |
4809 | |
4810 | if (!MEM_P (op0)) |
4811 | return NULL; |
4812 | |
4813 | op0 = XEXP (op0, 0); |
4814 | addrmode = GET_MODE (op0); |
4815 | if (addrmode == VOIDmode) |
4816 | addrmode = Pmode; |
4817 | |
4818 | op1 = expand_debug_expr (exp: offset); |
4819 | if (!op1) |
4820 | return NULL; |
4821 | |
4822 | offmode = GET_MODE (op1); |
4823 | if (offmode == VOIDmode) |
4824 | offmode = TYPE_MODE (TREE_TYPE (offset)); |
4825 | |
4826 | if (addrmode != offmode) |
4827 | op1 = lowpart_subreg (outermode: addrmode, op: op1, innermode: offmode); |
4828 | |
4829 | /* Don't use offset_address here, we don't need a |
4830 | recognizable address, and we don't want to generate |
4831 | code. */ |
4832 | op0 = gen_rtx_MEM (mode, simplify_gen_binary (code: PLUS, mode: addrmode, |
4833 | op0, op1)); |
4834 | } |
4835 | |
4836 | if (MEM_P (op0)) |
4837 | { |
4838 | if (mode1 == VOIDmode) |
4839 | { |
4840 | if (maybe_gt (bitsize, MAX_BITSIZE_MODE_ANY_INT)) |
4841 | return NULL; |
4842 | /* Bitfield. */ |
4843 | mode1 = smallest_int_mode_for_size (size: bitsize); |
4844 | } |
4845 | poly_int64 bytepos = bits_to_bytes_round_down (bitpos); |
4846 | if (maybe_ne (a: bytepos, b: 0)) |
4847 | { |
4848 | op0 = adjust_address_nv (op0, mode1, bytepos); |
4849 | bitpos = num_trailing_bits (bitpos); |
4850 | } |
4851 | else if (known_eq (bitpos, 0) |
4852 | && known_eq (bitsize, GET_MODE_BITSIZE (mode))) |
4853 | op0 = adjust_address_nv (op0, mode, 0); |
4854 | else if (GET_MODE (op0) != mode1) |
4855 | op0 = adjust_address_nv (op0, mode1, 0); |
4856 | else |
4857 | op0 = copy_rtx (op0); |
4858 | if (op0 == orig_op0) |
4859 | op0 = shallow_copy_rtx (op0); |
4860 | if (TREE_CODE (tem) != SSA_NAME) |
4861 | set_mem_attributes (op0, exp, 0); |
4862 | } |
4863 | |
4864 | if (known_eq (bitpos, 0) && mode == GET_MODE (op0)) |
4865 | return op0; |
4866 | |
4867 | if (maybe_lt (a: bitpos, b: 0)) |
4868 | return NULL; |
4869 | |
4870 | if (GET_MODE (op0) == BLKmode || mode == BLKmode) |
4871 | return NULL; |
4872 | |
4873 | poly_int64 bytepos; |
4874 | if (multiple_p (a: bitpos, BITS_PER_UNIT, multiple: &bytepos) |
4875 | && known_eq (bitsize, GET_MODE_BITSIZE (mode1))) |
4876 | { |
4877 | machine_mode opmode = GET_MODE (op0); |
4878 | |
4879 | if (opmode == VOIDmode) |
4880 | opmode = TYPE_MODE (TREE_TYPE (tem)); |
4881 | |
4882 | /* This condition may hold if we're expanding the address |
4883 | right past the end of an array that turned out not to |
4884 | be addressable (i.e., the address was only computed in |
4885 | debug stmts). The gen_subreg below would rightfully |
4886 | crash, and the address doesn't really exist, so just |
4887 | drop it. */ |
4888 | if (known_ge (bitpos, GET_MODE_BITSIZE (opmode))) |
4889 | return NULL; |
4890 | |
4891 | if (multiple_p (a: bitpos, b: GET_MODE_BITSIZE (mode))) |
4892 | return simplify_gen_subreg (outermode: mode, op: op0, innermode: opmode, byte: bytepos); |
4893 | } |
4894 | |
4895 | return simplify_gen_ternary (SCALAR_INT_MODE_P (GET_MODE (op0)) |
4896 | && TYPE_UNSIGNED (TREE_TYPE (exp)) |
4897 | ? SIGN_EXTRACT |
4898 | : ZERO_EXTRACT, mode, |
4899 | GET_MODE (op0) != VOIDmode |
4900 | ? GET_MODE (op0) |
4901 | : TYPE_MODE (TREE_TYPE (tem)), |
4902 | op0, op1: gen_int_mode (bitsize, word_mode), |
4903 | op2: gen_int_mode (bitpos, word_mode)); |
4904 | } |
4905 | |
4906 | case ABS_EXPR: |
4907 | case ABSU_EXPR: |
4908 | return simplify_gen_unary (code: ABS, mode, op: op0, op_mode: mode); |
4909 | |
4910 | case NEGATE_EXPR: |
4911 | return simplify_gen_unary (code: NEG, mode, op: op0, op_mode: mode); |
4912 | |
4913 | case BIT_NOT_EXPR: |
4914 | return simplify_gen_unary (code: NOT, mode, op: op0, op_mode: mode); |
4915 | |
4916 | case FLOAT_EXPR: |
4917 | return simplify_gen_unary (TYPE_UNSIGNED (TREE_TYPE (TREE_OPERAND (exp, |
4918 | 0))) |
4919 | ? UNSIGNED_FLOAT : FLOAT, mode, op: op0, |
4920 | op_mode: inner_mode); |
4921 | |
4922 | case FIX_TRUNC_EXPR: |
4923 | return simplify_gen_unary (code: unsignedp ? UNSIGNED_FIX : FIX, mode, op: op0, |
4924 | op_mode: inner_mode); |
4925 | |
4926 | case POINTER_PLUS_EXPR: |
4927 | /* For the rare target where pointers are not the same size as |
4928 | size_t, we need to check for mis-matched modes and correct |
4929 | the addend. */ |
4930 | if (op0 && op1 |
4931 | && is_a <scalar_int_mode> (GET_MODE (op0), result: &op0_mode) |
4932 | && is_a <scalar_int_mode> (GET_MODE (op1), result: &op1_mode) |
4933 | && op0_mode != op1_mode) |
4934 | { |
4935 | if (GET_MODE_BITSIZE (mode: op0_mode) < GET_MODE_BITSIZE (mode: op1_mode) |
4936 | /* If OP0 is a partial mode, then we must truncate, even |
4937 | if it has the same bitsize as OP1 as GCC's |
4938 | representation of partial modes is opaque. */ |
4939 | || (GET_MODE_CLASS (op0_mode) == MODE_PARTIAL_INT |
4940 | && (GET_MODE_BITSIZE (mode: op0_mode) |
4941 | == GET_MODE_BITSIZE (mode: op1_mode)))) |
4942 | op1 = simplify_gen_unary (code: TRUNCATE, mode: op0_mode, op: op1, op_mode: op1_mode); |
4943 | else |
4944 | /* We always sign-extend, regardless of the signedness of |
4945 | the operand, because the operand is always unsigned |
4946 | here even if the original C expression is signed. */ |
4947 | op1 = simplify_gen_unary (code: SIGN_EXTEND, mode: op0_mode, op: op1, op_mode: op1_mode); |
4948 | } |
4949 | /* Fall through. */ |
4950 | case PLUS_EXPR: |
4951 | return simplify_gen_binary (code: PLUS, mode, op0, op1); |
4952 | |
4953 | case MINUS_EXPR: |
4954 | case POINTER_DIFF_EXPR: |
4955 | return simplify_gen_binary (code: MINUS, mode, op0, op1); |
4956 | |
4957 | case MULT_EXPR: |
4958 | return simplify_gen_binary (code: MULT, mode, op0, op1); |
4959 | |
4960 | case RDIV_EXPR: |
4961 | case TRUNC_DIV_EXPR: |
4962 | case EXACT_DIV_EXPR: |
4963 | if (unsignedp) |
4964 | return simplify_gen_binary (code: UDIV, mode, op0, op1); |
4965 | else |
4966 | return simplify_gen_binary (code: DIV, mode, op0, op1); |
4967 | |
4968 | case TRUNC_MOD_EXPR: |
4969 | return simplify_gen_binary (code: unsignedp ? UMOD : MOD, mode, op0, op1); |
4970 | |
4971 | case FLOOR_DIV_EXPR: |
4972 | if (unsignedp) |
4973 | return simplify_gen_binary (code: UDIV, mode, op0, op1); |
4974 | else |
4975 | { |
4976 | rtx div = simplify_gen_binary (code: DIV, mode, op0, op1); |
4977 | rtx mod = simplify_gen_binary (code: MOD, mode, op0, op1); |
4978 | rtx adj = floor_sdiv_adjust (mode, mod, op1); |
4979 | return simplify_gen_binary (code: PLUS, mode, op0: div, op1: adj); |
4980 | } |
4981 | |
4982 | case FLOOR_MOD_EXPR: |
4983 | if (unsignedp) |
4984 | return simplify_gen_binary (code: UMOD, mode, op0, op1); |
4985 | else |
4986 | { |
4987 | rtx mod = simplify_gen_binary (code: MOD, mode, op0, op1); |
4988 | rtx adj = floor_sdiv_adjust (mode, mod, op1); |
4989 | adj = simplify_gen_unary (code: NEG, mode, |
4990 | op: simplify_gen_binary (code: MULT, mode, op0: adj, op1), |
4991 | op_mode: mode); |
4992 | return simplify_gen_binary (code: PLUS, mode, op0: mod, op1: adj); |
4993 | } |
4994 | |
4995 | case CEIL_DIV_EXPR: |
4996 | if (unsignedp) |
4997 | { |
4998 | rtx div = simplify_gen_binary (code: UDIV, mode, op0, op1); |
4999 | rtx mod = simplify_gen_binary (code: UMOD, mode, op0, op1); |
5000 | rtx adj = ceil_udiv_adjust (mode, mod, op1); |
5001 | return simplify_gen_binary (code: PLUS, mode, op0: div, op1: adj); |
5002 | } |
5003 | else |
5004 | { |
5005 | rtx div = simplify_gen_binary (code: DIV, mode, op0, op1); |
5006 | rtx mod = simplify_gen_binary (code: MOD, mode, op0, op1); |
5007 | rtx adj = ceil_sdiv_adjust (mode, mod, op1); |
5008 | return simplify_gen_binary (code: PLUS, mode, op0: div, op1: adj); |
5009 | } |
5010 | |
5011 | case CEIL_MOD_EXPR: |
5012 | if (unsignedp) |
5013 | { |
5014 | rtx mod = simplify_gen_binary (code: UMOD, mode, op0, op1); |
5015 | rtx adj = ceil_udiv_adjust (mode, mod, op1); |
5016 | adj = simplify_gen_unary (code: NEG, mode, |
5017 | op: simplify_gen_binary (code: MULT, mode, op0: adj, op1), |
5018 | op_mode: mode); |
5019 | return simplify_gen_binary (code: PLUS, mode, op0: mod, op1: adj); |
5020 | } |
5021 | else |
5022 | { |
5023 | rtx mod = simplify_gen_binary (code: MOD, mode, op0, op1); |
5024 | rtx adj = ceil_sdiv_adjust (mode, mod, op1); |
5025 | adj = simplify_gen_unary (code: NEG, mode, |
5026 | op: simplify_gen_binary (code: MULT, mode, op0: adj, op1), |
5027 | op_mode: mode); |
5028 | return simplify_gen_binary (code: PLUS, mode, op0: mod, op1: adj); |
5029 | } |
5030 | |
5031 | case ROUND_DIV_EXPR: |
5032 | if (unsignedp) |
5033 | { |
5034 | rtx div = simplify_gen_binary (code: UDIV, mode, op0, op1); |
5035 | rtx mod = simplify_gen_binary (code: UMOD, mode, op0, op1); |
5036 | rtx adj = round_udiv_adjust (mode, mod, op1); |
5037 | return simplify_gen_binary (code: PLUS, mode, op0: div, op1: adj); |
5038 | } |
5039 | else |
5040 | { |
5041 | rtx div = simplify_gen_binary (code: DIV, mode, op0, op1); |
5042 | rtx mod = simplify_gen_binary (code: MOD, mode, op0, op1); |
5043 | rtx adj = round_sdiv_adjust (mode, mod, op1); |
5044 | return simplify_gen_binary (code: PLUS, mode, op0: div, op1: adj); |
5045 | } |
5046 | |
5047 | case ROUND_MOD_EXPR: |
5048 | if (unsignedp) |
5049 | { |
5050 | rtx mod = simplify_gen_binary (code: UMOD, mode, op0, op1); |
5051 | rtx adj = round_udiv_adjust (mode, mod, op1); |
5052 | adj = simplify_gen_unary (code: NEG, mode, |
5053 | op: simplify_gen_binary (code: MULT, mode, op0: adj, op1), |
5054 | op_mode: mode); |
5055 | return simplify_gen_binary (code: PLUS, mode, op0: mod, op1: adj); |
5056 | } |
5057 | else |
5058 | { |
5059 | rtx mod = simplify_gen_binary (code: MOD, mode, op0, op1); |
5060 | rtx adj = round_sdiv_adjust (mode, mod, op1); |
5061 | adj = simplify_gen_unary (code: NEG, mode, |
5062 | op: simplify_gen_binary (code: MULT, mode, op0: adj, op1), |
5063 | op_mode: mode); |
5064 | return simplify_gen_binary (code: PLUS, mode, op0: mod, op1: adj); |
5065 | } |
5066 | |
5067 | case LSHIFT_EXPR: |
5068 | return simplify_gen_binary (code: ASHIFT, mode, op0, op1); |
5069 | |
5070 | case RSHIFT_EXPR: |
5071 | if (unsignedp) |
5072 | return simplify_gen_binary (code: LSHIFTRT, mode, op0, op1); |
5073 | else |
5074 | return simplify_gen_binary (code: ASHIFTRT, mode, op0, op1); |
5075 | |
5076 | case LROTATE_EXPR: |
5077 | return simplify_gen_binary (code: ROTATE, mode, op0, op1); |
5078 | |
5079 | case RROTATE_EXPR: |
5080 | return simplify_gen_binary (code: ROTATERT, mode, op0, op1); |
5081 | |
5082 | case MIN_EXPR: |
5083 | return simplify_gen_binary (code: unsignedp ? UMIN : SMIN, mode, op0, op1); |
5084 | |
5085 | case MAX_EXPR: |
5086 | return simplify_gen_binary (code: unsignedp ? UMAX : SMAX, mode, op0, op1); |
5087 | |
5088 | case BIT_AND_EXPR: |
5089 | case TRUTH_AND_EXPR: |
5090 | return simplify_gen_binary (code: AND, mode, op0, op1); |
5091 | |
5092 | case BIT_IOR_EXPR: |
5093 | case TRUTH_OR_EXPR: |
5094 | return simplify_gen_binary (code: IOR, mode, op0, op1); |
5095 | |
5096 | case BIT_XOR_EXPR: |
5097 | case TRUTH_XOR_EXPR: |
5098 | return simplify_gen_binary (code: XOR, mode, op0, op1); |
5099 | |
5100 | case TRUTH_ANDIF_EXPR: |
5101 | return gen_rtx_IF_THEN_ELSE (mode, op0, op1, const0_rtx); |
5102 | |
5103 | case TRUTH_ORIF_EXPR: |
5104 | return gen_rtx_IF_THEN_ELSE (mode, op0, const_true_rtx, op1); |
5105 | |
5106 | case TRUTH_NOT_EXPR: |
5107 | return simplify_gen_relational (code: EQ, mode, op_mode: inner_mode, op0, const0_rtx); |
5108 | |
5109 | case LT_EXPR: |
5110 | return simplify_gen_relational (code: unsignedp ? LTU : LT, mode, op_mode: inner_mode, |
5111 | op0, op1); |
5112 | |
5113 | case LE_EXPR: |
5114 | return simplify_gen_relational (code: unsignedp ? LEU : LE, mode, op_mode: inner_mode, |
5115 | op0, op1); |
5116 | |
5117 | case GT_EXPR: |
5118 | return simplify_gen_relational (code: unsignedp ? GTU : GT, mode, op_mode: inner_mode, |
5119 | op0, op1); |
5120 | |
5121 | case GE_EXPR: |
5122 | return simplify_gen_relational (code: unsignedp ? GEU : GE, mode, op_mode: inner_mode, |
5123 | op0, op1); |
5124 | |
5125 | case EQ_EXPR: |
5126 | return simplify_gen_relational (code: EQ, mode, op_mode: inner_mode, op0, op1); |
5127 | |
5128 | case NE_EXPR: |
5129 | return simplify_gen_relational (code: NE, mode, op_mode: inner_mode, op0, op1); |
5130 | |
5131 | case UNORDERED_EXPR: |
5132 | return simplify_gen_relational (code: UNORDERED, mode, op_mode: inner_mode, op0, op1); |
5133 | |
5134 | case ORDERED_EXPR: |
5135 | return simplify_gen_relational (code: ORDERED, mode, op_mode: inner_mode, op0, op1); |
5136 | |
5137 | case UNLT_EXPR: |
5138 | return simplify_gen_relational (code: UNLT, mode, op_mode: inner_mode, op0, op1); |
5139 | |
5140 | case UNLE_EXPR: |
5141 | return simplify_gen_relational (code: UNLE, mode, op_mode: inner_mode, op0, op1); |
5142 | |
5143 | case UNGT_EXPR: |
5144 | return simplify_gen_relational (code: UNGT, mode, op_mode: inner_mode, op0, op1); |
5145 | |
5146 | case UNGE_EXPR: |
5147 | return simplify_gen_relational (code: UNGE, mode, op_mode: inner_mode, op0, op1); |
5148 | |
5149 | case UNEQ_EXPR: |
5150 | return simplify_gen_relational (code: UNEQ, mode, op_mode: inner_mode, op0, op1); |
5151 | |
5152 | case LTGT_EXPR: |
5153 | return simplify_gen_relational (code: LTGT, mode, op_mode: inner_mode, op0, op1); |
5154 | |
5155 | case COND_EXPR: |
5156 | return gen_rtx_IF_THEN_ELSE (mode, op0, op1, op2); |
5157 | |
5158 | case COMPLEX_EXPR: |
5159 | gcc_assert (COMPLEX_MODE_P (mode)); |
5160 | if (GET_MODE (op0) == VOIDmode) |
5161 | op0 = gen_rtx_CONST (GET_MODE_INNER (mode), op0); |
5162 | if (GET_MODE (op1) == VOIDmode) |
5163 | op1 = gen_rtx_CONST (GET_MODE_INNER (mode), op1); |
5164 | return gen_rtx_CONCAT (mode, op0, op1); |
5165 | |
5166 | case CONJ_EXPR: |
5167 | if (GET_CODE (op0) == CONCAT) |
5168 | return gen_rtx_CONCAT (mode, XEXP (op0, 0), |
5169 | simplify_gen_unary (NEG, GET_MODE_INNER (mode), |
5170 | XEXP (op0, 1), |
5171 | GET_MODE_INNER (mode))); |
5172 | else |
5173 | { |
5174 | scalar_mode imode = GET_MODE_INNER (mode); |
5175 | rtx re, im; |
5176 | |
5177 | if (MEM_P (op0)) |
5178 | { |
5179 | re = adjust_address_nv (op0, imode, 0); |
5180 | im = adjust_address_nv (op0, imode, GET_MODE_SIZE (imode)); |
5181 | } |
5182 | else |
5183 | { |
5184 | scalar_int_mode ifmode; |
5185 | scalar_int_mode ihmode; |
5186 | rtx halfsize; |
5187 | if (!int_mode_for_mode (mode).exists (mode: &ifmode) |
5188 | || !int_mode_for_mode (imode).exists (mode: &ihmode)) |
5189 | return NULL; |
5190 | halfsize = GEN_INT (GET_MODE_BITSIZE (ihmode)); |
5191 | re = op0; |
5192 | if (mode != ifmode) |
5193 | re = gen_rtx_SUBREG (ifmode, re, 0); |
5194 | re = gen_rtx_ZERO_EXTRACT (ihmode, re, halfsize, const0_rtx); |
5195 | if (imode != ihmode) |
5196 | re = gen_rtx_SUBREG (imode, re, 0); |
5197 | im = copy_rtx (op0); |
5198 | if (mode != ifmode) |
5199 | im = gen_rtx_SUBREG (ifmode, im, 0); |
5200 | im = gen_rtx_ZERO_EXTRACT (ihmode, im, halfsize, halfsize); |
5201 | if (imode != ihmode) |
5202 | im = gen_rtx_SUBREG (imode, im, 0); |
5203 | } |
5204 | im = gen_rtx_NEG (imode, im); |
5205 | return gen_rtx_CONCAT (mode, re, im); |
5206 | } |
5207 | |
5208 | case ADDR_EXPR: |
5209 | op0 = expand_debug_expr (TREE_OPERAND (exp, 0)); |
5210 | if (!op0 || !MEM_P (op0)) |
5211 | { |
5212 | if ((TREE_CODE (TREE_OPERAND (exp, 0)) == VAR_DECL |
5213 | || TREE_CODE (TREE_OPERAND (exp, 0)) == PARM_DECL |
5214 | || TREE_CODE (TREE_OPERAND (exp, 0)) == RESULT_DECL) |
5215 | && (!TREE_ADDRESSABLE (TREE_OPERAND (exp, 0)) |
5216 | || target_for_debug_bind (TREE_OPERAND (exp, 0)))) |
5217 | return gen_rtx_DEBUG_IMPLICIT_PTR (mode, TREE_OPERAND (exp, 0)); |
5218 | |
5219 | if (handled_component_p (TREE_OPERAND (exp, 0))) |
5220 | { |
5221 | poly_int64 bitoffset, bitsize, maxsize, byteoffset; |
5222 | bool reverse; |
5223 | tree decl |
5224 | = get_ref_base_and_extent (TREE_OPERAND (exp, 0), &bitoffset, |
5225 | &bitsize, &maxsize, &reverse); |
5226 | if ((VAR_P (decl) |
5227 | || TREE_CODE (decl) == PARM_DECL |
5228 | || TREE_CODE (decl) == RESULT_DECL) |
5229 | && (!TREE_ADDRESSABLE (decl) |
5230 | || target_for_debug_bind (decl)) |
5231 | && multiple_p (a: bitoffset, BITS_PER_UNIT, multiple: &byteoffset) |
5232 | && known_gt (bitsize, 0) |
5233 | && known_eq (bitsize, maxsize)) |
5234 | { |
5235 | rtx base = gen_rtx_DEBUG_IMPLICIT_PTR (mode, decl); |
5236 | return plus_constant (mode, base, byteoffset); |
5237 | } |
5238 | } |
5239 | |
5240 | if (TREE_CODE (TREE_OPERAND (exp, 0)) == MEM_REF |
5241 | && TREE_CODE (TREE_OPERAND (TREE_OPERAND (exp, 0), 0)) |
5242 | == ADDR_EXPR) |
5243 | { |
5244 | op0 = expand_debug_expr (TREE_OPERAND (TREE_OPERAND (exp, 0), |
5245 | 0)); |
5246 | if (op0 != NULL |
5247 | && (GET_CODE (op0) == DEBUG_IMPLICIT_PTR |
5248 | || (GET_CODE (op0) == PLUS |
5249 | && GET_CODE (XEXP (op0, 0)) == DEBUG_IMPLICIT_PTR |
5250 | && CONST_INT_P (XEXP (op0, 1))))) |
5251 | { |
5252 | op1 = expand_debug_expr (TREE_OPERAND (TREE_OPERAND (exp, 0), |
5253 | 1)); |
5254 | poly_int64 offset; |
5255 | if (!op1 || !poly_int_rtx_p (x: op1, res: &offset)) |
5256 | return NULL; |
5257 | |
5258 | return plus_constant (mode, op0, offset); |
5259 | } |
5260 | } |
5261 | |
5262 | return NULL; |
5263 | } |
5264 | |
5265 | as = TYPE_ADDR_SPACE (TREE_TYPE (TREE_TYPE (exp))); |
5266 | addr_mode = SCALAR_INT_TYPE_MODE (TREE_TYPE (exp)); |
5267 | op0 = convert_debug_memory_address (mode: addr_mode, XEXP (op0, 0), as); |
5268 | |
5269 | return op0; |
5270 | |
5271 | case VECTOR_CST: |
5272 | { |
5273 | unsigned HOST_WIDE_INT i, nelts; |
5274 | |
5275 | if (!VECTOR_CST_NELTS (exp).is_constant (const_value: &nelts)) |
5276 | return NULL; |
5277 | |
5278 | op0 = gen_rtx_CONCATN (mode, rtvec_alloc (nelts)); |
5279 | |
5280 | for (i = 0; i < nelts; ++i) |
5281 | { |
5282 | op1 = expand_debug_expr (VECTOR_CST_ELT (exp, i)); |
5283 | if (!op1) |
5284 | return NULL; |
5285 | XVECEXP (op0, 0, i) = op1; |
5286 | } |
5287 | |
5288 | return op0; |
5289 | } |
5290 | |
5291 | case CONSTRUCTOR: |
5292 | if (TREE_CLOBBER_P (exp)) |
5293 | return NULL; |
5294 | else if (TREE_CODE (TREE_TYPE (exp)) == VECTOR_TYPE) |
5295 | { |
5296 | unsigned i; |
5297 | unsigned HOST_WIDE_INT nelts; |
5298 | tree val; |
5299 | |
5300 | if (!TYPE_VECTOR_SUBPARTS (TREE_TYPE (exp)).is_constant (const_value: &nelts)) |
5301 | goto flag_unsupported; |
5302 | |
5303 | op0 = gen_rtx_CONCATN (mode, rtvec_alloc (nelts)); |
5304 | |
5305 | FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (exp), i, val) |
5306 | { |
5307 | op1 = expand_debug_expr (exp: val); |
5308 | if (!op1) |
5309 | return NULL; |
5310 | XVECEXP (op0, 0, i) = op1; |
5311 | } |
5312 | |
5313 | if (i < nelts) |
5314 | { |
5315 | op1 = expand_debug_expr |
5316 | (exp: build_zero_cst (TREE_TYPE (TREE_TYPE (exp)))); |
5317 | |
5318 | if (!op1) |
5319 | return NULL; |
5320 | |
5321 | for (; i < nelts; i++) |
5322 | XVECEXP (op0, 0, i) = op1; |
5323 | } |
5324 | |
5325 | return op0; |
5326 | } |
5327 | else |
5328 | goto flag_unsupported; |
5329 | |
5330 | case CALL_EXPR: |
5331 | /* ??? Maybe handle some builtins? */ |
5332 | return NULL; |
5333 | |
5334 | case SSA_NAME: |
5335 | { |
5336 | gimple *g = get_gimple_for_ssa_name (exp); |
5337 | if (g) |
5338 | { |
5339 | tree t = NULL_TREE; |
5340 | if (deep_ter_debug_map) |
5341 | { |
5342 | tree *slot = deep_ter_debug_map->get (k: exp); |
5343 | if (slot) |
5344 | t = *slot; |
5345 | } |
5346 | if (t == NULL_TREE) |
5347 | t = gimple_assign_rhs_to_tree (stmt: g); |
5348 | op0 = expand_debug_expr (exp: t); |
5349 | if (!op0) |
5350 | return NULL; |
5351 | } |
5352 | else |
5353 | { |
5354 | /* If this is a reference to an incoming value of |
5355 | parameter that is never used in the code or where the |
5356 | incoming value is never used in the code, use |
5357 | PARM_DECL's DECL_RTL if set. */ |
5358 | if (SSA_NAME_IS_DEFAULT_DEF (exp) |
5359 | && SSA_NAME_VAR (exp) |
5360 | && TREE_CODE (SSA_NAME_VAR (exp)) == PARM_DECL |
5361 | && has_zero_uses (var: exp)) |
5362 | { |
5363 | op0 = expand_debug_parm_decl (SSA_NAME_VAR (exp)); |
5364 | if (op0) |
5365 | goto adjust_mode; |
5366 | op0 = expand_debug_expr (SSA_NAME_VAR (exp)); |
5367 | if (op0) |
5368 | goto adjust_mode; |
5369 | } |
5370 | |
5371 | int part = var_to_partition (map: SA.map, var: exp); |
5372 | |
5373 | if (part == NO_PARTITION) |
5374 | return NULL; |
5375 | |
5376 | gcc_assert (part >= 0 && (unsigned)part < SA.map->num_partitions); |
5377 | |
5378 | op0 = copy_rtx (SA.partition_to_pseudo[part]); |
5379 | } |
5380 | goto adjust_mode; |
5381 | } |
5382 | |
5383 | case ERROR_MARK: |
5384 | return NULL; |
5385 | |
5386 | /* Vector stuff. For most of the codes we don't have rtl codes. */ |
5387 | case REALIGN_LOAD_EXPR: |
5388 | case VEC_COND_EXPR: |
5389 | case VEC_PACK_FIX_TRUNC_EXPR: |
5390 | case VEC_PACK_FLOAT_EXPR: |
5391 | case VEC_PACK_SAT_EXPR: |
5392 | case VEC_PACK_TRUNC_EXPR: |
5393 | case VEC_UNPACK_FIX_TRUNC_HI_EXPR: |
5394 | case VEC_UNPACK_FIX_TRUNC_LO_EXPR: |
5395 | case VEC_UNPACK_FLOAT_HI_EXPR: |
5396 | case VEC_UNPACK_FLOAT_LO_EXPR: |
5397 | case VEC_UNPACK_HI_EXPR: |
5398 | case VEC_UNPACK_LO_EXPR: |
5399 | case VEC_WIDEN_MULT_HI_EXPR: |
5400 | case VEC_WIDEN_MULT_LO_EXPR: |
5401 | case VEC_WIDEN_MULT_EVEN_EXPR: |
5402 | case VEC_WIDEN_MULT_ODD_EXPR: |
5403 | case VEC_WIDEN_LSHIFT_HI_EXPR: |
5404 | case VEC_WIDEN_LSHIFT_LO_EXPR: |
5405 | case VEC_PERM_EXPR: |
5406 | case VEC_DUPLICATE_EXPR: |
5407 | case VEC_SERIES_EXPR: |
5408 | case SAD_EXPR: |
5409 | return NULL; |
5410 | |
5411 | /* Misc codes. */ |
5412 | case ADDR_SPACE_CONVERT_EXPR: |
5413 | case FIXED_CONVERT_EXPR: |
5414 | case OBJ_TYPE_REF: |
5415 | case WITH_SIZE_EXPR: |
5416 | case BIT_INSERT_EXPR: |
5417 | return NULL; |
5418 | |
5419 | case DOT_PROD_EXPR: |
5420 | if (SCALAR_INT_MODE_P (GET_MODE (op0)) |
5421 | && SCALAR_INT_MODE_P (mode)) |
5422 | { |
5423 | op0 |
5424 | = simplify_gen_unary (TYPE_UNSIGNED (TREE_TYPE (TREE_OPERAND (exp, |
5425 | 0))) |
5426 | ? ZERO_EXTEND : SIGN_EXTEND, mode, op: op0, |
5427 | op_mode: inner_mode); |
5428 | op1 |
5429 | = simplify_gen_unary (TYPE_UNSIGNED (TREE_TYPE (TREE_OPERAND (exp, |
5430 | 1))) |
5431 | ? ZERO_EXTEND : SIGN_EXTEND, mode, op: op1, |
5432 | op_mode: inner_mode); |
5433 | op0 = simplify_gen_binary (code: MULT, mode, op0, op1); |
5434 | return simplify_gen_binary (code: PLUS, mode, op0, op1: op2); |
5435 | } |
5436 | return NULL; |
5437 | |
5438 | case WIDEN_MULT_EXPR: |
5439 | case WIDEN_MULT_PLUS_EXPR: |
5440 | case WIDEN_MULT_MINUS_EXPR: |
5441 | if (SCALAR_INT_MODE_P (GET_MODE (op0)) |
5442 | && SCALAR_INT_MODE_P (mode)) |
5443 | { |
5444 | inner_mode = GET_MODE (op0); |
5445 | if (TYPE_UNSIGNED (TREE_TYPE (TREE_OPERAND (exp, 0)))) |
5446 | op0 = simplify_gen_unary (code: ZERO_EXTEND, mode, op: op0, op_mode: inner_mode); |
5447 | else |
5448 | op0 = simplify_gen_unary (code: SIGN_EXTEND, mode, op: op0, op_mode: inner_mode); |
5449 | if (TYPE_UNSIGNED (TREE_TYPE (TREE_OPERAND (exp, 1)))) |
5450 | op1 = simplify_gen_unary (code: ZERO_EXTEND, mode, op: op1, op_mode: inner_mode); |
5451 | else |
5452 | op1 = simplify_gen_unary (code: SIGN_EXTEND, mode, op: op1, op_mode: inner_mode); |
5453 | op0 = simplify_gen_binary (code: MULT, mode, op0, op1); |
5454 | if (TREE_CODE (exp) == WIDEN_MULT_EXPR) |
5455 | return op0; |
5456 | else if (TREE_CODE (exp) == WIDEN_MULT_PLUS_EXPR) |
5457 | return simplify_gen_binary (code: PLUS, mode, op0, op1: op2); |
5458 | else |
5459 | return simplify_gen_binary (code: MINUS, mode, op0: op2, op1: op0); |
5460 | } |
5461 | return NULL; |
5462 | |
5463 | case MULT_HIGHPART_EXPR: |
5464 | /* ??? Similar to the above. */ |
5465 | return NULL; |
5466 | |
5467 | case WIDEN_SUM_EXPR: |
5468 | case WIDEN_LSHIFT_EXPR: |
5469 | if (SCALAR_INT_MODE_P (GET_MODE (op0)) |
5470 | && SCALAR_INT_MODE_P (mode)) |
5471 | { |
5472 | op0 |
5473 | = simplify_gen_unary (TYPE_UNSIGNED (TREE_TYPE (TREE_OPERAND (exp, |
5474 | 0))) |
5475 | ? ZERO_EXTEND : SIGN_EXTEND, mode, op: op0, |
5476 | op_mode: inner_mode); |
5477 | return simplify_gen_binary (TREE_CODE (exp) == WIDEN_LSHIFT_EXPR |
5478 | ? ASHIFT : PLUS, mode, op0, op1); |
5479 | } |
5480 | return NULL; |
5481 | |
5482 | default: |
5483 | flag_unsupported: |
5484 | if (flag_checking) |
5485 | { |
5486 | debug_tree (exp); |
5487 | gcc_unreachable (); |
5488 | } |
5489 | return NULL; |
5490 | } |
5491 | } |
5492 | |
5493 | /* Return an RTX equivalent to the source bind value of the tree expression |
5494 | EXP. */ |
5495 | |
5496 | static rtx |
5497 | expand_debug_source_expr (tree exp) |
5498 | { |
5499 | rtx op0 = NULL_RTX; |
5500 | machine_mode mode = VOIDmode, inner_mode; |
5501 | |
5502 | switch (TREE_CODE (exp)) |
5503 | { |
5504 | case VAR_DECL: |
5505 | if (DECL_ABSTRACT_ORIGIN (exp)) |
5506 | return expand_debug_source_expr (DECL_ABSTRACT_ORIGIN (exp)); |
5507 | break; |
5508 | case PARM_DECL: |
5509 | { |
5510 | mode = DECL_MODE (exp); |
5511 | op0 = expand_debug_parm_decl (decl: exp); |
5512 | if (op0) |
5513 | break; |
5514 | /* See if this isn't an argument that has been completely |
5515 | optimized out. */ |
5516 | if (!DECL_RTL_SET_P (exp) |
5517 | && !DECL_INCOMING_RTL (exp) |
5518 | && DECL_ABSTRACT_ORIGIN (current_function_decl)) |
5519 | { |
5520 | tree aexp = DECL_ORIGIN (exp); |
5521 | if (DECL_CONTEXT (aexp) |
5522 | == DECL_ABSTRACT_ORIGIN (current_function_decl)) |
5523 | { |
5524 | vec<tree, va_gc> **debug_args; |
5525 | unsigned int ix; |
5526 | tree ddecl; |
5527 | debug_args = decl_debug_args_lookup (current_function_decl); |
5528 | if (debug_args != NULL) |
5529 | { |
5530 | for (ix = 0; vec_safe_iterate (v: *debug_args, ix, ptr: &ddecl); |
5531 | ix += 2) |
5532 | if (ddecl == aexp) |
5533 | return gen_rtx_DEBUG_PARAMETER_REF (mode, aexp); |
5534 | } |
5535 | } |
5536 | } |
5537 | break; |
5538 | } |
5539 | default: |
5540 | break; |
5541 | } |
5542 | |
5543 | if (op0 == NULL_RTX) |
5544 | return NULL_RTX; |
5545 | |
5546 | inner_mode = GET_MODE (op0); |
5547 | if (mode == inner_mode) |
5548 | return op0; |
5549 | |
5550 | if (FLOAT_MODE_P (mode) && FLOAT_MODE_P (inner_mode)) |
5551 | { |
5552 | if (GET_MODE_UNIT_BITSIZE (mode) |
5553 | == GET_MODE_UNIT_BITSIZE (inner_mode)) |
5554 | op0 = simplify_gen_subreg (outermode: mode, op: op0, innermode: inner_mode, byte: 0); |
5555 | else if (GET_MODE_UNIT_BITSIZE (mode) |
5556 | < GET_MODE_UNIT_BITSIZE (inner_mode)) |
5557 | op0 = simplify_gen_unary (code: FLOAT_TRUNCATE, mode, op: op0, op_mode: inner_mode); |
5558 | else |
5559 | op0 = simplify_gen_unary (code: FLOAT_EXTEND, mode, op: op0, op_mode: inner_mode); |
5560 | } |
5561 | else if (FLOAT_MODE_P (mode)) |
5562 | gcc_unreachable (); |
5563 | else if (FLOAT_MODE_P (inner_mode)) |
5564 | { |
5565 | if (TYPE_UNSIGNED (TREE_TYPE (exp))) |
5566 | op0 = simplify_gen_unary (code: UNSIGNED_FIX, mode, op: op0, op_mode: inner_mode); |
5567 | else |
5568 | op0 = simplify_gen_unary (code: FIX, mode, op: op0, op_mode: inner_mode); |
5569 | } |
5570 | else if (GET_MODE_UNIT_PRECISION (mode) |
5571 | == GET_MODE_UNIT_PRECISION (inner_mode)) |
5572 | op0 = lowpart_subreg (outermode: mode, op: op0, innermode: inner_mode); |
5573 | else if (GET_MODE_UNIT_PRECISION (mode) |
5574 | < GET_MODE_UNIT_PRECISION (inner_mode)) |
5575 | op0 = simplify_gen_unary (code: TRUNCATE, mode, op: op0, op_mode: inner_mode); |
5576 | else if (TYPE_UNSIGNED (TREE_TYPE (exp))) |
5577 | op0 = simplify_gen_unary (code: ZERO_EXTEND, mode, op: op0, op_mode: inner_mode); |
5578 | else |
5579 | op0 = simplify_gen_unary (code: SIGN_EXTEND, mode, op: op0, op_mode: inner_mode); |
5580 | |
5581 | return op0; |
5582 | } |
5583 | |
5584 | /* Ensure INSN_VAR_LOCATION_LOC (insn) doesn't have unbound complexity. |
5585 | Allow 4 levels of rtl nesting for most rtl codes, and if we see anything |
5586 | deeper than that, create DEBUG_EXPRs and emit DEBUG_INSNs before INSN. */ |
5587 | |
5588 | static void |
5589 | avoid_complex_debug_insns (rtx_insn *insn, rtx *exp_p, int depth) |
5590 | { |
5591 | rtx exp = *exp_p; |
5592 | |
5593 | if (exp == NULL_RTX) |
5594 | return; |
5595 | |
5596 | if ((OBJECT_P (exp) && !MEM_P (exp)) || GET_CODE (exp) == CLOBBER) |
5597 | return; |
5598 | |
5599 | if (depth == 4) |
5600 | { |
5601 | /* Create DEBUG_EXPR (and DEBUG_EXPR_DECL). */ |
5602 | rtx dval = make_debug_expr_from_rtl (exp); |
5603 | |
5604 | /* Emit a debug bind insn before INSN. */ |
5605 | rtx bind = gen_rtx_VAR_LOCATION (GET_MODE (exp), |
5606 | DEBUG_EXPR_TREE_DECL (dval), exp, |
5607 | VAR_INIT_STATUS_INITIALIZED); |
5608 | |
5609 | emit_debug_insn_before (bind, insn); |
5610 | *exp_p = dval; |
5611 | return; |
5612 | } |
5613 | |
5614 | const char *format_ptr = GET_RTX_FORMAT (GET_CODE (exp)); |
5615 | int i, j; |
5616 | for (i = 0; i < GET_RTX_LENGTH (GET_CODE (exp)); i++) |
5617 | switch (*format_ptr++) |
5618 | { |
5619 | case 'e': |
5620 | avoid_complex_debug_insns (insn, exp_p: &XEXP (exp, i), depth: depth + 1); |
5621 | break; |
5622 | |
5623 | case 'E': |
5624 | case 'V': |
5625 | for (j = 0; j < XVECLEN (exp, i); j++) |
5626 | avoid_complex_debug_insns (insn, exp_p: &XVECEXP (exp, i, j), depth: depth + 1); |
5627 | break; |
5628 | |
5629 | default: |
5630 | break; |
5631 | } |
5632 | } |
5633 | |
5634 | /* Expand the _LOCs in debug insns. We run this after expanding all |
5635 | regular insns, so that any variables referenced in the function |
5636 | will have their DECL_RTLs set. */ |
5637 | |
5638 | static void |
5639 | expand_debug_locations (void) |
5640 | { |
5641 | rtx_insn *insn; |
5642 | rtx_insn *last = get_last_insn (); |
5643 | int save_strict_alias = flag_strict_aliasing; |
5644 | |
5645 | /* New alias sets while setting up memory attributes cause |
5646 | -fcompare-debug failures, even though it doesn't bring about any |
5647 | codegen changes. */ |
5648 | flag_strict_aliasing = 0; |
5649 | |
5650 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
5651 | if (DEBUG_BIND_INSN_P (insn)) |
5652 | { |
5653 | tree value = (tree)INSN_VAR_LOCATION_LOC (insn); |
5654 | rtx val; |
5655 | rtx_insn *prev_insn, *insn2; |
5656 | machine_mode mode; |
5657 | |
5658 | if (value == NULL_TREE) |
5659 | val = NULL_RTX; |
5660 | else |
5661 | { |
5662 | if (INSN_VAR_LOCATION_STATUS (insn) |
5663 | == VAR_INIT_STATUS_UNINITIALIZED) |
5664 | val = expand_debug_source_expr (exp: value); |
5665 | /* The avoid_deep_ter_for_debug function inserts |
5666 | debug bind stmts after SSA_NAME definition, with the |
5667 | SSA_NAME as the whole bind location. Disable temporarily |
5668 | expansion of that SSA_NAME into the DEBUG_EXPR_DECL |
5669 | being defined in this DEBUG_INSN. */ |
5670 | else if (deep_ter_debug_map && TREE_CODE (value) == SSA_NAME) |
5671 | { |
5672 | tree *slot = deep_ter_debug_map->get (k: value); |
5673 | if (slot) |
5674 | { |
5675 | if (*slot == INSN_VAR_LOCATION_DECL (insn)) |
5676 | *slot = NULL_TREE; |
5677 | else |
5678 | slot = NULL; |
5679 | } |
5680 | val = expand_debug_expr (exp: value); |
5681 | if (slot) |
5682 | *slot = INSN_VAR_LOCATION_DECL (insn); |
5683 | } |
5684 | else |
5685 | val = expand_debug_expr (exp: value); |
5686 | gcc_assert (last == get_last_insn ()); |
5687 | } |
5688 | |
5689 | if (!val) |
5690 | val = gen_rtx_UNKNOWN_VAR_LOC (); |
5691 | else |
5692 | { |
5693 | mode = GET_MODE (INSN_VAR_LOCATION (insn)); |
5694 | |
5695 | gcc_assert (mode == GET_MODE (val) |
5696 | || (GET_MODE (val) == VOIDmode |
5697 | && (CONST_SCALAR_INT_P (val) |
5698 | || GET_CODE (val) == CONST_FIXED |
5699 | || GET_CODE (val) == LABEL_REF))); |
5700 | } |
5701 | |
5702 | INSN_VAR_LOCATION_LOC (insn) = val; |
5703 | prev_insn = PREV_INSN (insn); |
5704 | for (insn2 = insn; insn2 != prev_insn; insn2 = PREV_INSN (insn: insn2)) |
5705 | avoid_complex_debug_insns (insn: insn2, exp_p: &INSN_VAR_LOCATION_LOC (insn2), depth: 0); |
5706 | } |
5707 | |
5708 | flag_strict_aliasing = save_strict_alias; |
5709 | } |
5710 | |
5711 | /* Performs swapping operands of commutative operations to expand |
5712 | the expensive one first. */ |
5713 | |
5714 | static void |
5715 | reorder_operands (basic_block bb) |
5716 | { |
5717 | unsigned int *lattice; /* Hold cost of each statement. */ |
5718 | unsigned int i = 0, n = 0; |
5719 | gimple_stmt_iterator gsi; |
5720 | gimple_seq stmts; |
5721 | gimple *stmt; |
5722 | bool swap; |
5723 | tree op0, op1; |
5724 | ssa_op_iter iter; |
5725 | use_operand_p use_p; |
5726 | gimple *def0, *def1; |
5727 | |
5728 | /* Compute cost of each statement using estimate_num_insns. */ |
5729 | stmts = bb_seq (bb); |
5730 | for (gsi = gsi_start (seq&: stmts); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
5731 | { |
5732 | stmt = gsi_stmt (i: gsi); |
5733 | if (!is_gimple_debug (gs: stmt)) |
5734 | gimple_set_uid (g: stmt, uid: n++); |
5735 | } |
5736 | lattice = XNEWVEC (unsigned int, n); |
5737 | for (gsi = gsi_start (seq&: stmts); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
5738 | { |
5739 | unsigned cost; |
5740 | stmt = gsi_stmt (i: gsi); |
5741 | if (is_gimple_debug (gs: stmt)) |
5742 | continue; |
5743 | cost = estimate_num_insns (stmt, &eni_size_weights); |
5744 | lattice[i] = cost; |
5745 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) |
5746 | { |
5747 | tree use = USE_FROM_PTR (use_p); |
5748 | gimple *def_stmt; |
5749 | if (TREE_CODE (use) != SSA_NAME) |
5750 | continue; |
5751 | def_stmt = get_gimple_for_ssa_name (exp: use); |
5752 | if (!def_stmt) |
5753 | continue; |
5754 | lattice[i] += lattice[gimple_uid (g: def_stmt)]; |
5755 | } |
5756 | i++; |
5757 | if (!is_gimple_assign (gs: stmt) |
5758 | || !commutative_tree_code (gimple_assign_rhs_code (gs: stmt))) |
5759 | continue; |
5760 | op0 = gimple_op (gs: stmt, i: 1); |
5761 | op1 = gimple_op (gs: stmt, i: 2); |
5762 | if (TREE_CODE (op0) != SSA_NAME |
5763 | || TREE_CODE (op1) != SSA_NAME) |
5764 | continue; |
5765 | /* Swap operands if the second one is more expensive. */ |
5766 | def0 = get_gimple_for_ssa_name (exp: op0); |
5767 | def1 = get_gimple_for_ssa_name (exp: op1); |
5768 | if (!def1) |
5769 | continue; |
5770 | swap = false; |
5771 | if (!def0 || lattice[gimple_uid (g: def1)] > lattice[gimple_uid (g: def0)]) |
5772 | swap = true; |
5773 | if (swap) |
5774 | { |
5775 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5776 | { |
5777 | fprintf (stream: dump_file, format: "Swap operands in stmt:\n" ); |
5778 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
5779 | fprintf (stream: dump_file, format: "Cost left opnd=%d, right opnd=%d\n" , |
5780 | def0 ? lattice[gimple_uid (g: def0)] : 0, |
5781 | lattice[gimple_uid (g: def1)]); |
5782 | } |
5783 | swap_ssa_operands (stmt, gimple_assign_rhs1_ptr (gs: stmt), |
5784 | gimple_assign_rhs2_ptr (gs: stmt)); |
5785 | } |
5786 | } |
5787 | XDELETE (lattice); |
5788 | } |
5789 | |
5790 | /* Expand basic block BB from GIMPLE trees to RTL. */ |
5791 | |
5792 | static basic_block |
5793 | expand_gimple_basic_block (basic_block bb, bool disable_tail_calls) |
5794 | { |
5795 | gimple_stmt_iterator gsi; |
5796 | gimple_seq stmts; |
5797 | gimple *stmt = NULL; |
5798 | rtx_note *note = NULL; |
5799 | rtx_insn *last; |
5800 | edge e; |
5801 | edge_iterator ei; |
5802 | bool nondebug_stmt_seen = false; |
5803 | |
5804 | if (dump_file) |
5805 | fprintf (stream: dump_file, format: "\n;; Generating RTL for gimple basic block %d\n" , |
5806 | bb->index); |
5807 | |
5808 | /* Note that since we are now transitioning from GIMPLE to RTL, we |
5809 | cannot use the gsi_*_bb() routines because they expect the basic |
5810 | block to be in GIMPLE, instead of RTL. Therefore, we need to |
5811 | access the BB sequence directly. */ |
5812 | if (optimize) |
5813 | reorder_operands (bb); |
5814 | stmts = bb_seq (bb); |
5815 | bb->il.gimple.seq = NULL; |
5816 | bb->il.gimple.phi_nodes = NULL; |
5817 | rtl_profile_for_bb (bb); |
5818 | init_rtl_bb_info (bb); |
5819 | bb->flags |= BB_RTL; |
5820 | |
5821 | /* Remove the RETURN_EXPR if we may fall though to the exit |
5822 | instead. */ |
5823 | gsi = gsi_last (seq&: stmts); |
5824 | if (!gsi_end_p (i: gsi) |
5825 | && gimple_code (g: gsi_stmt (i: gsi)) == GIMPLE_RETURN) |
5826 | { |
5827 | greturn *ret_stmt = as_a <greturn *> (p: gsi_stmt (i: gsi)); |
5828 | |
5829 | gcc_assert (single_succ_p (bb)); |
5830 | gcc_assert (single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun)); |
5831 | |
5832 | if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) |
5833 | && !gimple_return_retval (gs: ret_stmt)) |
5834 | { |
5835 | gsi_remove (&gsi, false); |
5836 | single_succ_edge (bb)->flags |= EDGE_FALLTHRU; |
5837 | } |
5838 | } |
5839 | |
5840 | gsi = gsi_start (seq&: stmts); |
5841 | if (!gsi_end_p (i: gsi)) |
5842 | { |
5843 | stmt = gsi_stmt (i: gsi); |
5844 | if (gimple_code (g: stmt) != GIMPLE_LABEL) |
5845 | stmt = NULL; |
5846 | } |
5847 | |
5848 | rtx_code_label **elt = lab_rtx_for_bb->get (k: bb); |
5849 | |
5850 | if (stmt || elt) |
5851 | { |
5852 | gcc_checking_assert (!note); |
5853 | last = get_last_insn (); |
5854 | |
5855 | if (stmt) |
5856 | { |
5857 | expand_gimple_stmt (stmt); |
5858 | gsi_next (i: &gsi); |
5859 | } |
5860 | |
5861 | if (elt) |
5862 | emit_label (*elt); |
5863 | |
5864 | BB_HEAD (bb) = NEXT_INSN (insn: last); |
5865 | if (NOTE_P (BB_HEAD (bb))) |
5866 | BB_HEAD (bb) = NEXT_INSN (BB_HEAD (bb)); |
5867 | gcc_assert (LABEL_P (BB_HEAD (bb))); |
5868 | note = emit_note_after (NOTE_INSN_BASIC_BLOCK, BB_HEAD (bb)); |
5869 | |
5870 | maybe_dump_rtl_for_gimple_stmt (stmt, since: last); |
5871 | } |
5872 | else |
5873 | BB_HEAD (bb) = note = emit_note (NOTE_INSN_BASIC_BLOCK); |
5874 | |
5875 | if (note) |
5876 | NOTE_BASIC_BLOCK (note) = bb; |
5877 | |
5878 | for (; !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
5879 | { |
5880 | basic_block new_bb; |
5881 | |
5882 | stmt = gsi_stmt (i: gsi); |
5883 | if (!is_gimple_debug (gs: stmt)) |
5884 | nondebug_stmt_seen = true; |
5885 | |
5886 | /* If this statement is a non-debug one, and we generate debug |
5887 | insns, then this one might be the last real use of a TERed |
5888 | SSA_NAME, but where there are still some debug uses further |
5889 | down. Expanding the current SSA name in such further debug |
5890 | uses by their RHS might lead to wrong debug info, as coalescing |
5891 | might make the operands of such RHS be placed into the same |
5892 | pseudo as something else. Like so: |
5893 | a_1 = a_0 + 1; // Assume a_1 is TERed and a_0 is dead |
5894 | use(a_1); |
5895 | a_2 = ... |
5896 | #DEBUG ... => a_1 |
5897 | As a_0 and a_2 don't overlap in lifetime, assume they are coalesced. |
5898 | If we now would expand a_1 by it's RHS (a_0 + 1) in the debug use, |
5899 | the write to a_2 would actually have clobbered the place which |
5900 | formerly held a_0. |
5901 | |
5902 | So, instead of that, we recognize the situation, and generate |
5903 | debug temporaries at the last real use of TERed SSA names: |
5904 | a_1 = a_0 + 1; |
5905 | #DEBUG #D1 => a_1 |
5906 | use(a_1); |
5907 | a_2 = ... |
5908 | #DEBUG ... => #D1 |
5909 | */ |
5910 | if (MAY_HAVE_DEBUG_BIND_INSNS |
5911 | && SA.values |
5912 | && !is_gimple_debug (gs: stmt)) |
5913 | { |
5914 | ssa_op_iter iter; |
5915 | tree op; |
5916 | gimple *def; |
5917 | |
5918 | location_t sloc = curr_insn_location (); |
5919 | |
5920 | /* Look for SSA names that have their last use here (TERed |
5921 | names always have only one real use). */ |
5922 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) |
5923 | if ((def = get_gimple_for_ssa_name (exp: op))) |
5924 | { |
5925 | imm_use_iterator imm_iter; |
5926 | use_operand_p use_p; |
5927 | bool have_debug_uses = false; |
5928 | |
5929 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, op) |
5930 | { |
5931 | if (gimple_debug_bind_p (USE_STMT (use_p))) |
5932 | { |
5933 | have_debug_uses = true; |
5934 | break; |
5935 | } |
5936 | } |
5937 | |
5938 | if (have_debug_uses) |
5939 | { |
5940 | /* OP is a TERed SSA name, with DEF its defining |
5941 | statement, and where OP is used in further debug |
5942 | instructions. Generate a debug temporary, and |
5943 | replace all uses of OP in debug insns with that |
5944 | temporary. */ |
5945 | gimple *debugstmt; |
5946 | tree value = gimple_assign_rhs_to_tree (stmt: def); |
5947 | tree vexpr = build_debug_expr_decl (TREE_TYPE (value)); |
5948 | rtx val; |
5949 | machine_mode mode; |
5950 | |
5951 | set_curr_insn_location (gimple_location (g: def)); |
5952 | |
5953 | if (DECL_P (value)) |
5954 | mode = DECL_MODE (value); |
5955 | else |
5956 | mode = TYPE_MODE (TREE_TYPE (value)); |
5957 | /* FIXME: Is setting the mode really necessary? */ |
5958 | SET_DECL_MODE (vexpr, mode); |
5959 | |
5960 | val = gen_rtx_VAR_LOCATION |
5961 | (mode, vexpr, (rtx)value, VAR_INIT_STATUS_INITIALIZED); |
5962 | |
5963 | emit_debug_insn (val); |
5964 | |
5965 | FOR_EACH_IMM_USE_STMT (debugstmt, imm_iter, op) |
5966 | { |
5967 | if (!gimple_debug_bind_p (s: debugstmt)) |
5968 | continue; |
5969 | |
5970 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) |
5971 | SET_USE (use_p, vexpr); |
5972 | |
5973 | update_stmt (s: debugstmt); |
5974 | } |
5975 | } |
5976 | } |
5977 | set_curr_insn_location (sloc); |
5978 | } |
5979 | |
5980 | currently_expanding_gimple_stmt = stmt; |
5981 | |
5982 | /* Expand this statement, then evaluate the resulting RTL and |
5983 | fixup the CFG accordingly. */ |
5984 | if (gimple_code (g: stmt) == GIMPLE_COND) |
5985 | { |
5986 | new_bb = expand_gimple_cond (bb, stmt: as_a <gcond *> (p: stmt)); |
5987 | if (new_bb) |
5988 | { |
5989 | currently_expanding_gimple_stmt = NULL; |
5990 | return new_bb; |
5991 | } |
5992 | } |
5993 | else if (is_gimple_debug (gs: stmt)) |
5994 | { |
5995 | location_t sloc = curr_insn_location (); |
5996 | gimple_stmt_iterator nsi = gsi; |
5997 | |
5998 | for (;;) |
5999 | { |
6000 | tree var; |
6001 | tree value = NULL_TREE; |
6002 | rtx val = NULL_RTX; |
6003 | machine_mode mode; |
6004 | |
6005 | if (!gimple_debug_nonbind_marker_p (s: stmt)) |
6006 | { |
6007 | if (gimple_debug_bind_p (s: stmt)) |
6008 | { |
6009 | var = gimple_debug_bind_get_var (dbg: stmt); |
6010 | |
6011 | if (TREE_CODE (var) != DEBUG_EXPR_DECL |
6012 | && TREE_CODE (var) != LABEL_DECL |
6013 | && !target_for_debug_bind (var)) |
6014 | goto delink_debug_stmt; |
6015 | |
6016 | if (DECL_P (var) && !VECTOR_TYPE_P (TREE_TYPE (var))) |
6017 | mode = DECL_MODE (var); |
6018 | else |
6019 | mode = TYPE_MODE (TREE_TYPE (var)); |
6020 | |
6021 | if (gimple_debug_bind_has_value_p (dbg: stmt)) |
6022 | value = gimple_debug_bind_get_value (dbg: stmt); |
6023 | |
6024 | val = gen_rtx_VAR_LOCATION |
6025 | (mode, var, (rtx)value, VAR_INIT_STATUS_INITIALIZED); |
6026 | } |
6027 | else if (gimple_debug_source_bind_p (s: stmt)) |
6028 | { |
6029 | var = gimple_debug_source_bind_get_var (dbg: stmt); |
6030 | |
6031 | value = gimple_debug_source_bind_get_value (dbg: stmt); |
6032 | |
6033 | if (!VECTOR_TYPE_P (TREE_TYPE (var))) |
6034 | mode = DECL_MODE (var); |
6035 | else |
6036 | mode = TYPE_MODE (TREE_TYPE (var)); |
6037 | |
6038 | val = gen_rtx_VAR_LOCATION (mode, var, (rtx)value, |
6039 | VAR_INIT_STATUS_UNINITIALIZED); |
6040 | } |
6041 | else |
6042 | gcc_unreachable (); |
6043 | } |
6044 | /* If this function was first compiled with markers |
6045 | enabled, but they're now disable (e.g. LTO), drop |
6046 | them on the floor. */ |
6047 | else if (gimple_debug_nonbind_marker_p (s: stmt) |
6048 | && !MAY_HAVE_DEBUG_MARKER_INSNS) |
6049 | goto delink_debug_stmt; |
6050 | else if (gimple_debug_begin_stmt_p (s: stmt)) |
6051 | val = GEN_RTX_DEBUG_MARKER_BEGIN_STMT_PAT (); |
6052 | else if (gimple_debug_inline_entry_p (s: stmt)) |
6053 | val = GEN_RTX_DEBUG_MARKER_INLINE_ENTRY_PAT (); |
6054 | else |
6055 | gcc_unreachable (); |
6056 | |
6057 | last = get_last_insn (); |
6058 | |
6059 | set_curr_insn_location (gimple_location (g: stmt)); |
6060 | |
6061 | emit_debug_insn (val); |
6062 | |
6063 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6064 | { |
6065 | /* We can't dump the insn with a TREE where an RTX |
6066 | is expected. */ |
6067 | if (GET_CODE (val) == VAR_LOCATION) |
6068 | { |
6069 | gcc_checking_assert (PAT_VAR_LOCATION_LOC (val) == (rtx)value); |
6070 | PAT_VAR_LOCATION_LOC (val) = const0_rtx; |
6071 | } |
6072 | maybe_dump_rtl_for_gimple_stmt (stmt, since: last); |
6073 | if (GET_CODE (val) == VAR_LOCATION) |
6074 | PAT_VAR_LOCATION_LOC (val) = (rtx)value; |
6075 | } |
6076 | |
6077 | delink_debug_stmt: |
6078 | /* In order not to generate too many debug temporaries, |
6079 | we delink all uses of debug statements we already expanded. |
6080 | Therefore debug statements between definition and real |
6081 | use of TERed SSA names will continue to use the SSA name, |
6082 | and not be replaced with debug temps. */ |
6083 | delink_stmt_imm_use (stmt); |
6084 | |
6085 | gsi = nsi; |
6086 | gsi_next (i: &nsi); |
6087 | if (gsi_end_p (i: nsi)) |
6088 | break; |
6089 | stmt = gsi_stmt (i: nsi); |
6090 | if (!is_gimple_debug (gs: stmt)) |
6091 | break; |
6092 | } |
6093 | |
6094 | set_curr_insn_location (sloc); |
6095 | } |
6096 | else |
6097 | { |
6098 | gcall *call_stmt = dyn_cast <gcall *> (p: stmt); |
6099 | if (call_stmt |
6100 | && gimple_call_tail_p (s: call_stmt) |
6101 | && disable_tail_calls) |
6102 | gimple_call_set_tail (s: call_stmt, tail_p: false); |
6103 | |
6104 | if (call_stmt && gimple_call_tail_p (s: call_stmt)) |
6105 | { |
6106 | bool can_fallthru; |
6107 | new_bb = expand_gimple_tailcall (bb, stmt: call_stmt, can_fallthru: &can_fallthru); |
6108 | if (new_bb) |
6109 | { |
6110 | if (can_fallthru) |
6111 | bb = new_bb; |
6112 | else |
6113 | { |
6114 | currently_expanding_gimple_stmt = NULL; |
6115 | return new_bb; |
6116 | } |
6117 | } |
6118 | } |
6119 | else |
6120 | { |
6121 | def_operand_p def_p; |
6122 | def_p = SINGLE_SSA_DEF_OPERAND (stmt, SSA_OP_DEF); |
6123 | |
6124 | if (def_p != NULL) |
6125 | { |
6126 | /* Ignore this stmt if it is in the list of |
6127 | replaceable expressions. */ |
6128 | if (SA.values |
6129 | && bitmap_bit_p (SA.values, |
6130 | SSA_NAME_VERSION (DEF_FROM_PTR (def_p)))) |
6131 | continue; |
6132 | } |
6133 | last = expand_gimple_stmt (stmt); |
6134 | maybe_dump_rtl_for_gimple_stmt (stmt, since: last); |
6135 | } |
6136 | } |
6137 | } |
6138 | |
6139 | currently_expanding_gimple_stmt = NULL; |
6140 | |
6141 | /* Expand implicit goto and convert goto_locus. */ |
6142 | FOR_EACH_EDGE (e, ei, bb->succs) |
6143 | { |
6144 | if (e->goto_locus != UNKNOWN_LOCATION || !nondebug_stmt_seen) |
6145 | set_curr_insn_location (e->goto_locus); |
6146 | if ((e->flags & EDGE_FALLTHRU) && e->dest != bb->next_bb) |
6147 | { |
6148 | emit_jump (label_rtx_for_bb (bb: e->dest)); |
6149 | e->flags &= ~EDGE_FALLTHRU; |
6150 | } |
6151 | } |
6152 | |
6153 | /* Expanded RTL can create a jump in the last instruction of block. |
6154 | This later might be assumed to be a jump to successor and break edge insertion. |
6155 | We need to insert dummy move to prevent this. PR41440. */ |
6156 | if (single_succ_p (bb) |
6157 | && (single_succ_edge (bb)->flags & EDGE_FALLTHRU) |
6158 | && (last = get_last_insn ()) |
6159 | && (JUMP_P (last) |
6160 | || (DEBUG_INSN_P (last) |
6161 | && JUMP_P (prev_nondebug_insn (last))))) |
6162 | { |
6163 | rtx dummy = gen_reg_rtx (SImode); |
6164 | emit_insn_after_noloc (gen_move_insn (dummy, dummy), last, NULL); |
6165 | } |
6166 | |
6167 | do_pending_stack_adjust (); |
6168 | |
6169 | /* Find the block tail. The last insn in the block is the insn |
6170 | before a barrier and/or table jump insn. */ |
6171 | last = get_last_insn (); |
6172 | if (BARRIER_P (last)) |
6173 | last = PREV_INSN (insn: last); |
6174 | if (JUMP_TABLE_DATA_P (last)) |
6175 | last = PREV_INSN (insn: PREV_INSN (insn: last)); |
6176 | if (BARRIER_P (last)) |
6177 | last = PREV_INSN (insn: last); |
6178 | BB_END (bb) = last; |
6179 | |
6180 | update_bb_for_insn (bb); |
6181 | |
6182 | return bb; |
6183 | } |
6184 | |
6185 | |
6186 | /* Create a basic block for initialization code. */ |
6187 | |
6188 | static basic_block |
6189 | construct_init_block (void) |
6190 | { |
6191 | basic_block init_block, first_block; |
6192 | edge e = NULL; |
6193 | int flags; |
6194 | |
6195 | /* Multiple entry points not supported yet. */ |
6196 | gcc_assert (EDGE_COUNT (ENTRY_BLOCK_PTR_FOR_FN (cfun)->succs) == 1); |
6197 | init_rtl_bb_info (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
6198 | init_rtl_bb_info (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
6199 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->flags |= BB_RTL; |
6200 | EXIT_BLOCK_PTR_FOR_FN (cfun)->flags |= BB_RTL; |
6201 | |
6202 | e = EDGE_SUCC (ENTRY_BLOCK_PTR_FOR_FN (cfun), 0); |
6203 | |
6204 | /* When entry edge points to first basic block, we don't need jump, |
6205 | otherwise we have to jump into proper target. */ |
6206 | if (e && e->dest != ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb) |
6207 | { |
6208 | tree label = gimple_block_label (e->dest); |
6209 | |
6210 | emit_jump (jump_target_rtx (label)); |
6211 | flags = 0; |
6212 | } |
6213 | else |
6214 | flags = EDGE_FALLTHRU; |
6215 | |
6216 | init_block = create_basic_block (NEXT_INSN (insn: get_insns ()), |
6217 | get_last_insn (), |
6218 | ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
6219 | init_block->count = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count; |
6220 | add_bb_to_loop (init_block, ENTRY_BLOCK_PTR_FOR_FN (cfun)->loop_father); |
6221 | if (e) |
6222 | { |
6223 | first_block = e->dest; |
6224 | redirect_edge_succ (e, init_block); |
6225 | make_single_succ_edge (init_block, first_block, flags); |
6226 | } |
6227 | else |
6228 | make_single_succ_edge (init_block, EXIT_BLOCK_PTR_FOR_FN (cfun), |
6229 | EDGE_FALLTHRU); |
6230 | |
6231 | update_bb_for_insn (init_block); |
6232 | return init_block; |
6233 | } |
6234 | |
6235 | /* For each lexical block, set BLOCK_NUMBER to the depth at which it is |
6236 | found in the block tree. */ |
6237 | |
6238 | static void |
6239 | set_block_levels (tree block, int level) |
6240 | { |
6241 | while (block) |
6242 | { |
6243 | BLOCK_NUMBER (block) = level; |
6244 | set_block_levels (BLOCK_SUBBLOCKS (block), level: level + 1); |
6245 | block = BLOCK_CHAIN (block); |
6246 | } |
6247 | } |
6248 | |
6249 | /* Create a block containing landing pads and similar stuff. */ |
6250 | |
6251 | static void |
6252 | construct_exit_block (void) |
6253 | { |
6254 | rtx_insn *head = get_last_insn (); |
6255 | rtx_insn *end; |
6256 | basic_block exit_block; |
6257 | edge e, e2; |
6258 | unsigned ix; |
6259 | edge_iterator ei; |
6260 | basic_block prev_bb = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb; |
6261 | rtx_insn *orig_end = BB_END (prev_bb); |
6262 | |
6263 | rtl_profile_for_bb (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
6264 | |
6265 | /* Make sure the locus is set to the end of the function, so that |
6266 | epilogue line numbers and warnings are set properly. */ |
6267 | if (LOCATION_LOCUS (cfun->function_end_locus) != UNKNOWN_LOCATION) |
6268 | input_location = cfun->function_end_locus; |
6269 | |
6270 | /* Generate rtl for function exit. */ |
6271 | expand_function_end (); |
6272 | |
6273 | end = get_last_insn (); |
6274 | if (head == end) |
6275 | return; |
6276 | /* While emitting the function end we could move end of the last basic |
6277 | block. */ |
6278 | BB_END (prev_bb) = orig_end; |
6279 | while (NEXT_INSN (insn: head) && NOTE_P (NEXT_INSN (head))) |
6280 | head = NEXT_INSN (insn: head); |
6281 | /* But make sure exit_block starts with RETURN_LABEL, otherwise the |
6282 | bb count counting will be confused. Any instructions before that |
6283 | label are emitted for the case where PREV_BB falls through into the |
6284 | exit block, so append those instructions to prev_bb in that case. */ |
6285 | if (NEXT_INSN (insn: head) != return_label) |
6286 | { |
6287 | while (NEXT_INSN (insn: head) != return_label) |
6288 | { |
6289 | if (!NOTE_P (NEXT_INSN (head))) |
6290 | BB_END (prev_bb) = NEXT_INSN (insn: head); |
6291 | head = NEXT_INSN (insn: head); |
6292 | } |
6293 | } |
6294 | exit_block = create_basic_block (NEXT_INSN (insn: head), end, prev_bb); |
6295 | exit_block->count = EXIT_BLOCK_PTR_FOR_FN (cfun)->count; |
6296 | add_bb_to_loop (exit_block, EXIT_BLOCK_PTR_FOR_FN (cfun)->loop_father); |
6297 | |
6298 | ix = 0; |
6299 | while (ix < EDGE_COUNT (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)) |
6300 | { |
6301 | e = EDGE_PRED (EXIT_BLOCK_PTR_FOR_FN (cfun), ix); |
6302 | if (!(e->flags & EDGE_ABNORMAL)) |
6303 | redirect_edge_succ (e, exit_block); |
6304 | else |
6305 | ix++; |
6306 | } |
6307 | |
6308 | e = make_single_succ_edge (exit_block, EXIT_BLOCK_PTR_FOR_FN (cfun), |
6309 | EDGE_FALLTHRU); |
6310 | FOR_EACH_EDGE (e2, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) |
6311 | if (e2 != e) |
6312 | { |
6313 | exit_block->count -= e2->count (); |
6314 | } |
6315 | update_bb_for_insn (exit_block); |
6316 | } |
6317 | |
6318 | /* Helper function for discover_nonconstant_array_refs. |
6319 | Look for ARRAY_REF nodes with non-constant indexes and mark them |
6320 | addressable. */ |
6321 | |
6322 | static tree |
6323 | discover_nonconstant_array_refs_r (tree * tp, int *walk_subtrees, |
6324 | void *data) |
6325 | { |
6326 | tree t = *tp; |
6327 | bitmap forced_stack_vars = (bitmap)((walk_stmt_info *)data)->info; |
6328 | |
6329 | if (IS_TYPE_OR_DECL_P (t)) |
6330 | *walk_subtrees = 0; |
6331 | else if (REFERENCE_CLASS_P (t) && TREE_THIS_VOLATILE (t)) |
6332 | { |
6333 | t = get_base_address (t); |
6334 | if (t && DECL_P (t) |
6335 | && DECL_MODE (t) != BLKmode |
6336 | && !TREE_ADDRESSABLE (t)) |
6337 | bitmap_set_bit (forced_stack_vars, DECL_UID (t)); |
6338 | *walk_subtrees = 0; |
6339 | } |
6340 | else if (TREE_CODE (t) == ARRAY_REF || TREE_CODE (t) == ARRAY_RANGE_REF) |
6341 | { |
6342 | while (((TREE_CODE (t) == ARRAY_REF || TREE_CODE (t) == ARRAY_RANGE_REF) |
6343 | && is_gimple_min_invariant (TREE_OPERAND (t, 1)) |
6344 | && (!TREE_OPERAND (t, 2) |
6345 | || is_gimple_min_invariant (TREE_OPERAND (t, 2)))) |
6346 | || (TREE_CODE (t) == COMPONENT_REF |
6347 | && (!TREE_OPERAND (t,2) |
6348 | || is_gimple_min_invariant (TREE_OPERAND (t, 2)))) |
6349 | || TREE_CODE (t) == BIT_FIELD_REF |
6350 | || TREE_CODE (t) == REALPART_EXPR |
6351 | || TREE_CODE (t) == IMAGPART_EXPR |
6352 | || TREE_CODE (t) == VIEW_CONVERT_EXPR |
6353 | || CONVERT_EXPR_P (t)) |
6354 | t = TREE_OPERAND (t, 0); |
6355 | |
6356 | if (TREE_CODE (t) == ARRAY_REF || TREE_CODE (t) == ARRAY_RANGE_REF) |
6357 | { |
6358 | t = get_base_address (t); |
6359 | if (t && DECL_P (t) |
6360 | && DECL_MODE (t) != BLKmode |
6361 | && !TREE_ADDRESSABLE (t)) |
6362 | bitmap_set_bit (forced_stack_vars, DECL_UID (t)); |
6363 | } |
6364 | |
6365 | *walk_subtrees = 0; |
6366 | } |
6367 | /* References of size POLY_INT_CST to a fixed-size object must go |
6368 | through memory. It's more efficient to force that here than |
6369 | to create temporary slots on the fly. |
6370 | RTL expansion expectes TARGET_MEM_REF to always address actual memory. |
6371 | Also, force to stack non-BLKmode vars accessed through VIEW_CONVERT_EXPR |
6372 | to BLKmode type. */ |
6373 | else if (TREE_CODE (t) == TARGET_MEM_REF |
6374 | || (TREE_CODE (t) == MEM_REF |
6375 | && TYPE_SIZE (TREE_TYPE (t)) |
6376 | && POLY_INT_CST_P (TYPE_SIZE (TREE_TYPE (t)))) |
6377 | || (TREE_CODE (t) == VIEW_CONVERT_EXPR |
6378 | && TYPE_MODE (TREE_TYPE (t)) == BLKmode)) |
6379 | { |
6380 | tree base = get_base_address (t); |
6381 | if (base |
6382 | && DECL_P (base) |
6383 | && !TREE_ADDRESSABLE (base) |
6384 | && DECL_MODE (base) != BLKmode |
6385 | && GET_MODE_SIZE (DECL_MODE (base)).is_constant ()) |
6386 | bitmap_set_bit (forced_stack_vars, DECL_UID (base)); |
6387 | *walk_subtrees = 0; |
6388 | } |
6389 | |
6390 | return NULL_TREE; |
6391 | } |
6392 | |
6393 | /* If there's a chance to get a pseudo for t then if it would be of float mode |
6394 | and the actual access is via an integer mode (lowered memcpy or similar |
6395 | access) then avoid the register expansion if the mode likely is not storage |
6396 | suitable for raw bits processing (like XFmode on i?86). */ |
6397 | |
6398 | static void |
6399 | avoid_type_punning_on_regs (tree t, bitmap forced_stack_vars) |
6400 | { |
6401 | machine_mode access_mode = TYPE_MODE (TREE_TYPE (t)); |
6402 | if (access_mode != BLKmode |
6403 | && !SCALAR_INT_MODE_P (access_mode)) |
6404 | return; |
6405 | tree base = get_base_address (t); |
6406 | if (DECL_P (base) |
6407 | && !TREE_ADDRESSABLE (base) |
6408 | && FLOAT_MODE_P (DECL_MODE (base)) |
6409 | && maybe_lt (a: GET_MODE_PRECISION (DECL_MODE (base)), |
6410 | b: GET_MODE_BITSIZE (GET_MODE_INNER (DECL_MODE (base)))) |
6411 | /* Double check in the expensive way we really would get a pseudo. */ |
6412 | && use_register_for_decl (base)) |
6413 | bitmap_set_bit (forced_stack_vars, DECL_UID (base)); |
6414 | } |
6415 | |
6416 | /* RTL expansion is not able to compile array references with variable |
6417 | offsets for arrays stored in single register. Discover such |
6418 | expressions and mark variables as addressable to avoid this |
6419 | scenario. */ |
6420 | |
6421 | static void |
6422 | discover_nonconstant_array_refs (bitmap forced_stack_vars) |
6423 | { |
6424 | basic_block bb; |
6425 | gimple_stmt_iterator gsi; |
6426 | |
6427 | walk_stmt_info wi = {}; |
6428 | wi.info = forced_stack_vars; |
6429 | FOR_EACH_BB_FN (bb, cfun) |
6430 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
6431 | { |
6432 | gimple *stmt = gsi_stmt (i: gsi); |
6433 | if (!is_gimple_debug (gs: stmt)) |
6434 | { |
6435 | walk_gimple_op (stmt, discover_nonconstant_array_refs_r, &wi); |
6436 | gcall *call = dyn_cast <gcall *> (p: stmt); |
6437 | if (call && gimple_call_internal_p (gs: call)) |
6438 | { |
6439 | tree cand = NULL_TREE; |
6440 | switch (gimple_call_internal_fn (gs: call)) |
6441 | { |
6442 | case IFN_LOAD_LANES: |
6443 | /* The source must be a MEM. */ |
6444 | cand = gimple_call_arg (gs: call, index: 0); |
6445 | break; |
6446 | case IFN_STORE_LANES: |
6447 | /* The destination must be a MEM. */ |
6448 | cand = gimple_call_lhs (gs: call); |
6449 | break; |
6450 | default: |
6451 | break; |
6452 | } |
6453 | if (cand) |
6454 | cand = get_base_address (t: cand); |
6455 | if (cand |
6456 | && DECL_P (cand) |
6457 | && use_register_for_decl (cand)) |
6458 | bitmap_set_bit (forced_stack_vars, DECL_UID (cand)); |
6459 | } |
6460 | if (gimple_vdef (g: stmt)) |
6461 | { |
6462 | tree t = gimple_get_lhs (stmt); |
6463 | if (t && REFERENCE_CLASS_P (t)) |
6464 | avoid_type_punning_on_regs (t, forced_stack_vars); |
6465 | } |
6466 | } |
6467 | } |
6468 | } |
6469 | |
6470 | /* This function sets crtl->args.internal_arg_pointer to a virtual |
6471 | register if DRAP is needed. Local register allocator will replace |
6472 | virtual_incoming_args_rtx with the virtual register. */ |
6473 | |
6474 | static void |
6475 | expand_stack_alignment (void) |
6476 | { |
6477 | rtx drap_rtx; |
6478 | unsigned int preferred_stack_boundary; |
6479 | |
6480 | if (! SUPPORTS_STACK_ALIGNMENT) |
6481 | return; |
6482 | |
6483 | if (cfun->calls_alloca |
6484 | || cfun->has_nonlocal_label |
6485 | || crtl->has_nonlocal_goto) |
6486 | crtl->need_drap = true; |
6487 | |
6488 | /* Call update_stack_boundary here again to update incoming stack |
6489 | boundary. It may set incoming stack alignment to a different |
6490 | value after RTL expansion. TARGET_FUNCTION_OK_FOR_SIBCALL may |
6491 | use the minimum incoming stack alignment to check if it is OK |
6492 | to perform sibcall optimization since sibcall optimization will |
6493 | only align the outgoing stack to incoming stack boundary. */ |
6494 | if (targetm.calls.update_stack_boundary) |
6495 | targetm.calls.update_stack_boundary (); |
6496 | |
6497 | /* The incoming stack frame has to be aligned at least at |
6498 | parm_stack_boundary. */ |
6499 | gcc_assert (crtl->parm_stack_boundary <= INCOMING_STACK_BOUNDARY); |
6500 | |
6501 | /* Update crtl->stack_alignment_estimated and use it later to align |
6502 | stack. We check PREFERRED_STACK_BOUNDARY if there may be non-call |
6503 | exceptions since callgraph doesn't collect incoming stack alignment |
6504 | in this case. */ |
6505 | if (cfun->can_throw_non_call_exceptions |
6506 | && PREFERRED_STACK_BOUNDARY > crtl->preferred_stack_boundary) |
6507 | preferred_stack_boundary = PREFERRED_STACK_BOUNDARY; |
6508 | else |
6509 | preferred_stack_boundary = crtl->preferred_stack_boundary; |
6510 | if (preferred_stack_boundary > crtl->stack_alignment_estimated) |
6511 | crtl->stack_alignment_estimated = preferred_stack_boundary; |
6512 | if (preferred_stack_boundary > crtl->stack_alignment_needed) |
6513 | crtl->stack_alignment_needed = preferred_stack_boundary; |
6514 | |
6515 | gcc_assert (crtl->stack_alignment_needed |
6516 | <= crtl->stack_alignment_estimated); |
6517 | |
6518 | crtl->stack_realign_needed |
6519 | = INCOMING_STACK_BOUNDARY < crtl->stack_alignment_estimated; |
6520 | crtl->stack_realign_tried = crtl->stack_realign_needed; |
6521 | |
6522 | crtl->stack_realign_processed = true; |
6523 | |
6524 | /* Target has to redefine TARGET_GET_DRAP_RTX to support stack |
6525 | alignment. */ |
6526 | gcc_assert (targetm.calls.get_drap_rtx != NULL); |
6527 | drap_rtx = targetm.calls.get_drap_rtx (); |
6528 | |
6529 | /* stack_realign_drap and drap_rtx must match. */ |
6530 | gcc_assert ((stack_realign_drap != 0) == (drap_rtx != NULL)); |
6531 | |
6532 | /* Do nothing if NULL is returned, which means DRAP is not needed. */ |
6533 | if (drap_rtx != NULL) |
6534 | { |
6535 | crtl->args.internal_arg_pointer = drap_rtx; |
6536 | |
6537 | /* Call fixup_tail_calls to clean up REG_EQUIV note if DRAP is |
6538 | needed. */ |
6539 | fixup_tail_calls (); |
6540 | } |
6541 | } |
6542 | |
6543 | |
6544 | static void |
6545 | expand_main_function (void) |
6546 | { |
6547 | #if (defined(INVOKE__main) \ |
6548 | || (!defined(HAS_INIT_SECTION) \ |
6549 | && !defined(INIT_SECTION_ASM_OP) \ |
6550 | && !defined(INIT_ARRAY_SECTION_ASM_OP))) |
6551 | emit_library_call (init_one_libfunc (NAME__MAIN), LCT_NORMAL, VOIDmode); |
6552 | #endif |
6553 | } |
6554 | |
6555 | |
6556 | /* Expand code to initialize the stack_protect_guard. This is invoked at |
6557 | the beginning of a function to be protected. */ |
6558 | |
6559 | static void |
6560 | stack_protect_prologue (void) |
6561 | { |
6562 | tree guard_decl = targetm.stack_protect_guard (); |
6563 | rtx x, y; |
6564 | |
6565 | crtl->stack_protect_guard_decl = guard_decl; |
6566 | x = expand_normal (crtl->stack_protect_guard); |
6567 | |
6568 | if (targetm.have_stack_protect_combined_set () && guard_decl) |
6569 | { |
6570 | gcc_assert (DECL_P (guard_decl)); |
6571 | y = DECL_RTL (guard_decl); |
6572 | |
6573 | /* Allow the target to compute address of Y and copy it to X without |
6574 | leaking Y into a register. This combined address + copy pattern |
6575 | allows the target to prevent spilling of any intermediate results by |
6576 | splitting it after register allocator. */ |
6577 | if (rtx_insn *insn = targetm.gen_stack_protect_combined_set (x, y)) |
6578 | { |
6579 | emit_insn (insn); |
6580 | return; |
6581 | } |
6582 | } |
6583 | |
6584 | if (guard_decl) |
6585 | y = expand_normal (exp: guard_decl); |
6586 | else |
6587 | y = const0_rtx; |
6588 | |
6589 | /* Allow the target to copy from Y to X without leaking Y into a |
6590 | register. */ |
6591 | if (targetm.have_stack_protect_set ()) |
6592 | if (rtx_insn *insn = targetm.gen_stack_protect_set (x, y)) |
6593 | { |
6594 | emit_insn (insn); |
6595 | return; |
6596 | } |
6597 | |
6598 | /* Otherwise do a straight move. */ |
6599 | emit_move_insn (x, y); |
6600 | } |
6601 | |
6602 | /* Translate the intermediate representation contained in the CFG |
6603 | from GIMPLE trees to RTL. |
6604 | |
6605 | We do conversion per basic block and preserve/update the tree CFG. |
6606 | This implies we have to do some magic as the CFG can simultaneously |
6607 | consist of basic blocks containing RTL and GIMPLE trees. This can |
6608 | confuse the CFG hooks, so be careful to not manipulate CFG during |
6609 | the expansion. */ |
6610 | |
6611 | namespace { |
6612 | |
6613 | const pass_data pass_data_expand = |
6614 | { |
6615 | .type: RTL_PASS, /* type */ |
6616 | .name: "expand" , /* name */ |
6617 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
6618 | .tv_id: TV_EXPAND, /* tv_id */ |
6619 | .properties_required: ( PROP_ssa | PROP_gimple_leh | PROP_cfg |
6620 | | PROP_gimple_lcx |
6621 | | PROP_gimple_lvec |
6622 | | PROP_gimple_lva), /* properties_required */ |
6623 | PROP_rtl, /* properties_provided */ |
6624 | .properties_destroyed: ( PROP_ssa | PROP_gimple ), /* properties_destroyed */ |
6625 | .todo_flags_start: 0, /* todo_flags_start */ |
6626 | .todo_flags_finish: 0, /* todo_flags_finish */ |
6627 | }; |
6628 | |
6629 | class pass_expand : public rtl_opt_pass |
6630 | { |
6631 | public: |
6632 | pass_expand (gcc::context *ctxt) |
6633 | : rtl_opt_pass (pass_data_expand, ctxt) |
6634 | {} |
6635 | |
6636 | /* opt_pass methods: */ |
6637 | unsigned int execute (function *) final override; |
6638 | |
6639 | }; // class pass_expand |
6640 | |
6641 | unsigned int |
6642 | pass_expand::execute (function *fun) |
6643 | { |
6644 | basic_block bb, init_block; |
6645 | edge_iterator ei; |
6646 | edge e; |
6647 | rtx_insn *var_seq, *var_ret_seq; |
6648 | unsigned i; |
6649 | |
6650 | timevar_push (tv: TV_OUT_OF_SSA); |
6651 | rewrite_out_of_ssa (sa: &SA); |
6652 | timevar_pop (tv: TV_OUT_OF_SSA); |
6653 | SA.partition_to_pseudo = XCNEWVEC (rtx, SA.map->num_partitions); |
6654 | |
6655 | if (MAY_HAVE_DEBUG_BIND_STMTS && flag_tree_ter) |
6656 | { |
6657 | gimple_stmt_iterator gsi; |
6658 | FOR_EACH_BB_FN (bb, cfun) |
6659 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
6660 | if (gimple_debug_bind_p (s: gsi_stmt (i: gsi))) |
6661 | avoid_deep_ter_for_debug (stmt: gsi_stmt (i: gsi), depth: 0); |
6662 | } |
6663 | |
6664 | /* Mark arrays indexed with non-constant indices with TREE_ADDRESSABLE. */ |
6665 | auto_bitmap forced_stack_vars; |
6666 | discover_nonconstant_array_refs (forced_stack_vars); |
6667 | |
6668 | /* Make sure all values used by the optimization passes have sane |
6669 | defaults. */ |
6670 | reg_renumber = 0; |
6671 | |
6672 | /* Some backends want to know that we are expanding to RTL. */ |
6673 | currently_expanding_to_rtl = 1; |
6674 | /* Dominators are not kept up-to-date as we may create new basic-blocks. */ |
6675 | free_dominance_info (CDI_DOMINATORS); |
6676 | |
6677 | rtl_profile_for_bb (ENTRY_BLOCK_PTR_FOR_FN (fun)); |
6678 | |
6679 | insn_locations_init (); |
6680 | if (!DECL_IS_UNDECLARED_BUILTIN (current_function_decl)) |
6681 | { |
6682 | /* Eventually, all FEs should explicitly set function_start_locus. */ |
6683 | if (LOCATION_LOCUS (fun->function_start_locus) == UNKNOWN_LOCATION) |
6684 | set_curr_insn_location |
6685 | (DECL_SOURCE_LOCATION (current_function_decl)); |
6686 | else |
6687 | set_curr_insn_location (fun->function_start_locus); |
6688 | } |
6689 | else |
6690 | set_curr_insn_location (UNKNOWN_LOCATION); |
6691 | prologue_location = curr_insn_location (); |
6692 | |
6693 | #ifdef INSN_SCHEDULING |
6694 | init_sched_attrs (); |
6695 | #endif |
6696 | |
6697 | /* Make sure first insn is a note even if we don't want linenums. |
6698 | This makes sure the first insn will never be deleted. |
6699 | Also, final expects a note to appear there. */ |
6700 | emit_note (NOTE_INSN_DELETED); |
6701 | |
6702 | targetm.expand_to_rtl_hook (); |
6703 | crtl->init_stack_alignment (); |
6704 | fun->cfg->max_jumptable_ents = 0; |
6705 | |
6706 | /* Resovle the function section. Some targets, like ARM EABI rely on knowledge |
6707 | of the function section at exapnsion time to predict distance of calls. */ |
6708 | resolve_unique_section (current_function_decl, 0, flag_function_sections); |
6709 | |
6710 | /* Expand the variables recorded during gimple lowering. */ |
6711 | timevar_push (tv: TV_VAR_EXPAND); |
6712 | start_sequence (); |
6713 | |
6714 | var_ret_seq = expand_used_vars (forced_stack_vars); |
6715 | |
6716 | var_seq = get_insns (); |
6717 | end_sequence (); |
6718 | timevar_pop (tv: TV_VAR_EXPAND); |
6719 | |
6720 | /* Honor stack protection warnings. */ |
6721 | if (warn_stack_protect) |
6722 | { |
6723 | if (fun->calls_alloca) |
6724 | warning (OPT_Wstack_protector, |
6725 | "stack protector not protecting local variables: " |
6726 | "variable length buffer" ); |
6727 | if (has_short_buffer && !crtl->stack_protect_guard) |
6728 | warning (OPT_Wstack_protector, |
6729 | "stack protector not protecting function: " |
6730 | "all local arrays are less than %d bytes long" , |
6731 | (int) param_ssp_buffer_size); |
6732 | } |
6733 | |
6734 | /* Temporarily mark PARM_DECLs and RESULT_DECLs we need to expand to |
6735 | memory addressable so expand_function_start can emit the required |
6736 | copies. */ |
6737 | auto_vec<tree, 16> marked_parms; |
6738 | for (tree parm = DECL_ARGUMENTS (current_function_decl); parm; |
6739 | parm = DECL_CHAIN (parm)) |
6740 | if (!TREE_ADDRESSABLE (parm) |
6741 | && bitmap_bit_p (forced_stack_vars, DECL_UID (parm))) |
6742 | { |
6743 | TREE_ADDRESSABLE (parm) = 1; |
6744 | marked_parms.safe_push (obj: parm); |
6745 | } |
6746 | if (DECL_RESULT (current_function_decl) |
6747 | && !TREE_ADDRESSABLE (DECL_RESULT (current_function_decl)) |
6748 | && bitmap_bit_p (forced_stack_vars, |
6749 | DECL_UID (DECL_RESULT (current_function_decl)))) |
6750 | { |
6751 | TREE_ADDRESSABLE (DECL_RESULT (current_function_decl)) = 1; |
6752 | marked_parms.safe_push (DECL_RESULT (current_function_decl)); |
6753 | } |
6754 | |
6755 | /* Set up parameters and prepare for return, for the function. */ |
6756 | expand_function_start (current_function_decl); |
6757 | |
6758 | /* Clear TREE_ADDRESSABLE again. */ |
6759 | while (!marked_parms.is_empty ()) |
6760 | TREE_ADDRESSABLE (marked_parms.pop ()) = 0; |
6761 | |
6762 | /* If we emitted any instructions for setting up the variables, |
6763 | emit them before the FUNCTION_START note. */ |
6764 | if (var_seq) |
6765 | { |
6766 | emit_insn_before (var_seq, parm_birth_insn); |
6767 | |
6768 | /* In expand_function_end we'll insert the alloca save/restore |
6769 | before parm_birth_insn. We've just insertted an alloca call. |
6770 | Adjust the pointer to match. */ |
6771 | parm_birth_insn = var_seq; |
6772 | } |
6773 | |
6774 | /* Now propagate the RTL assignment of each partition to the |
6775 | underlying var of each SSA_NAME. */ |
6776 | tree name; |
6777 | |
6778 | FOR_EACH_SSA_NAME (i, name, cfun) |
6779 | { |
6780 | /* We might have generated new SSA names in |
6781 | update_alias_info_with_stack_vars. They will have a NULL |
6782 | defining statements, and won't be part of the partitioning, |
6783 | so ignore those. */ |
6784 | if (!SSA_NAME_DEF_STMT (name)) |
6785 | continue; |
6786 | |
6787 | adjust_one_expanded_partition_var (var: name); |
6788 | } |
6789 | |
6790 | /* Clean up RTL of variables that straddle across multiple |
6791 | partitions, and check that the rtl of any PARM_DECLs that are not |
6792 | cleaned up is that of their default defs. */ |
6793 | FOR_EACH_SSA_NAME (i, name, cfun) |
6794 | { |
6795 | int part; |
6796 | |
6797 | /* We might have generated new SSA names in |
6798 | update_alias_info_with_stack_vars. They will have a NULL |
6799 | defining statements, and won't be part of the partitioning, |
6800 | so ignore those. */ |
6801 | if (!SSA_NAME_DEF_STMT (name)) |
6802 | continue; |
6803 | part = var_to_partition (map: SA.map, var: name); |
6804 | if (part == NO_PARTITION) |
6805 | continue; |
6806 | |
6807 | /* If this decl was marked as living in multiple places, reset |
6808 | this now to NULL. */ |
6809 | tree var = SSA_NAME_VAR (name); |
6810 | if (var && DECL_RTL_IF_SET (var) == pc_rtx) |
6811 | SET_DECL_RTL (var, NULL); |
6812 | /* Check that the pseudos chosen by assign_parms are those of |
6813 | the corresponding default defs. */ |
6814 | else if (SSA_NAME_IS_DEFAULT_DEF (name) |
6815 | && (TREE_CODE (var) == PARM_DECL |
6816 | || TREE_CODE (var) == RESULT_DECL)) |
6817 | { |
6818 | rtx in = DECL_RTL_IF_SET (var); |
6819 | gcc_assert (in); |
6820 | rtx out = SA.partition_to_pseudo[part]; |
6821 | gcc_assert (in == out); |
6822 | |
6823 | /* Now reset VAR's RTL to IN, so that the _EXPR attrs match |
6824 | those expected by debug backends for each parm and for |
6825 | the result. This is particularly important for stabs, |
6826 | whose register elimination from parm's DECL_RTL may cause |
6827 | -fcompare-debug differences as SET_DECL_RTL changes reg's |
6828 | attrs. So, make sure the RTL already has the parm as the |
6829 | EXPR, so that it won't change. */ |
6830 | SET_DECL_RTL (var, NULL_RTX); |
6831 | if (MEM_P (in)) |
6832 | set_mem_attributes (in, var, true); |
6833 | SET_DECL_RTL (var, in); |
6834 | } |
6835 | } |
6836 | |
6837 | /* If this function is `main', emit a call to `__main' |
6838 | to run global initializers, etc. */ |
6839 | if (DECL_NAME (current_function_decl) |
6840 | && MAIN_NAME_P (DECL_NAME (current_function_decl)) |
6841 | && DECL_FILE_SCOPE_P (current_function_decl)) |
6842 | expand_main_function (); |
6843 | |
6844 | /* Initialize the stack_protect_guard field. This must happen after the |
6845 | call to __main (if any) so that the external decl is initialized. */ |
6846 | if (crtl->stack_protect_guard && targetm.stack_protect_runtime_enabled_p ()) |
6847 | stack_protect_prologue (); |
6848 | |
6849 | expand_phi_nodes (sa: &SA); |
6850 | |
6851 | /* Release any stale SSA redirection data. */ |
6852 | redirect_edge_var_map_empty (); |
6853 | |
6854 | /* Register rtl specific functions for cfg. */ |
6855 | rtl_register_cfg_hooks (); |
6856 | |
6857 | init_block = construct_init_block (); |
6858 | |
6859 | /* Clear EDGE_EXECUTABLE on the entry edge(s). It is cleaned from the |
6860 | remaining edges later. */ |
6861 | FOR_EACH_EDGE (e, ei, ENTRY_BLOCK_PTR_FOR_FN (fun)->succs) |
6862 | e->flags &= ~EDGE_EXECUTABLE; |
6863 | |
6864 | /* If the function has too many markers, drop them while expanding. */ |
6865 | if (cfun->debug_marker_count |
6866 | >= param_max_debug_marker_count) |
6867 | cfun->debug_nonbind_markers = false; |
6868 | |
6869 | lab_rtx_for_bb = new hash_map<basic_block, rtx_code_label *>; |
6870 | FOR_BB_BETWEEN (bb, init_block->next_bb, EXIT_BLOCK_PTR_FOR_FN (fun), |
6871 | next_bb) |
6872 | bb = expand_gimple_basic_block (bb, disable_tail_calls: var_ret_seq != NULL_RTX); |
6873 | |
6874 | if (MAY_HAVE_DEBUG_BIND_INSNS) |
6875 | expand_debug_locations (); |
6876 | |
6877 | if (deep_ter_debug_map) |
6878 | { |
6879 | delete deep_ter_debug_map; |
6880 | deep_ter_debug_map = NULL; |
6881 | } |
6882 | |
6883 | /* Free stuff we no longer need after GIMPLE optimizations. */ |
6884 | free_dominance_info (CDI_DOMINATORS); |
6885 | free_dominance_info (CDI_POST_DOMINATORS); |
6886 | delete_tree_cfg_annotations (fun); |
6887 | |
6888 | timevar_push (tv: TV_OUT_OF_SSA); |
6889 | finish_out_of_ssa (sa: &SA); |
6890 | timevar_pop (tv: TV_OUT_OF_SSA); |
6891 | |
6892 | timevar_push (tv: TV_POST_EXPAND); |
6893 | /* We are no longer in SSA form. */ |
6894 | fun->gimple_df->in_ssa_p = false; |
6895 | loops_state_clear (flags: LOOP_CLOSED_SSA); |
6896 | |
6897 | /* Expansion is used by optimization passes too, set maybe_hot_insn_p |
6898 | conservatively to true until they are all profile aware. */ |
6899 | delete lab_rtx_for_bb; |
6900 | free_histograms (fun); |
6901 | |
6902 | construct_exit_block (); |
6903 | insn_locations_finalize (); |
6904 | |
6905 | if (var_ret_seq) |
6906 | { |
6907 | rtx_insn *after = return_label; |
6908 | rtx_insn *next = NEXT_INSN (insn: after); |
6909 | if (next && NOTE_INSN_BASIC_BLOCK_P (next)) |
6910 | after = next; |
6911 | emit_insn_after (var_ret_seq, after); |
6912 | } |
6913 | |
6914 | if (hwasan_sanitize_stack_p ()) |
6915 | hwasan_maybe_emit_frame_base_init (); |
6916 | |
6917 | /* Zap the tree EH table. */ |
6918 | set_eh_throw_stmt_table (fun, NULL); |
6919 | |
6920 | /* We need JUMP_LABEL be set in order to redirect jumps, and hence |
6921 | split edges which edge insertions might do. */ |
6922 | rebuild_jump_labels (get_insns ()); |
6923 | |
6924 | /* If we have a single successor to the entry block, put the pending insns |
6925 | after parm birth, but before NOTE_INSNS_FUNCTION_BEG. */ |
6926 | if (single_succ_p (ENTRY_BLOCK_PTR_FOR_FN (fun))) |
6927 | { |
6928 | edge e = single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (fun)); |
6929 | if (e->insns.r) |
6930 | { |
6931 | rtx_insn *insns = e->insns.r; |
6932 | e->insns.r = NULL; |
6933 | rebuild_jump_labels_chain (insns); |
6934 | if (NOTE_P (parm_birth_insn) |
6935 | && NOTE_KIND (parm_birth_insn) == NOTE_INSN_FUNCTION_BEG) |
6936 | emit_insn_before_noloc (insns, parm_birth_insn, e->dest); |
6937 | else |
6938 | emit_insn_after_noloc (insns, parm_birth_insn, e->dest); |
6939 | } |
6940 | } |
6941 | |
6942 | /* Otherwise, as well as for other edges, take the usual way. */ |
6943 | commit_edge_insertions (); |
6944 | |
6945 | /* We're done expanding trees to RTL. */ |
6946 | currently_expanding_to_rtl = 0; |
6947 | |
6948 | flush_mark_addressable_queue (); |
6949 | |
6950 | FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (fun)->next_bb, |
6951 | EXIT_BLOCK_PTR_FOR_FN (fun), next_bb) |
6952 | { |
6953 | edge e; |
6954 | edge_iterator ei; |
6955 | for (ei = ei_start (bb->succs); (e = ei_safe_edge (i: ei)); ) |
6956 | { |
6957 | /* Clear EDGE_EXECUTABLE. This flag is never used in the backend. */ |
6958 | e->flags &= ~EDGE_EXECUTABLE; |
6959 | |
6960 | /* At the moment not all abnormal edges match the RTL |
6961 | representation. It is safe to remove them here as |
6962 | find_many_sub_basic_blocks will rediscover them. |
6963 | In the future we should get this fixed properly. */ |
6964 | if ((e->flags & EDGE_ABNORMAL) |
6965 | && !(e->flags & EDGE_SIBCALL)) |
6966 | remove_edge (e); |
6967 | else |
6968 | ei_next (i: &ei); |
6969 | } |
6970 | } |
6971 | |
6972 | auto_sbitmap blocks (last_basic_block_for_fn (fun)); |
6973 | bitmap_ones (blocks); |
6974 | find_many_sub_basic_blocks (blocks); |
6975 | purge_all_dead_edges (); |
6976 | |
6977 | /* After initial rtl generation, call back to finish generating |
6978 | exception support code. We need to do this before cleaning up |
6979 | the CFG as the code does not expect dead landing pads. */ |
6980 | if (fun->eh->region_tree != NULL) |
6981 | finish_eh_generation (); |
6982 | |
6983 | /* Call expand_stack_alignment after finishing all |
6984 | updates to crtl->preferred_stack_boundary. */ |
6985 | expand_stack_alignment (); |
6986 | |
6987 | /* Fixup REG_EQUIV notes in the prologue if there are tailcalls in this |
6988 | function. */ |
6989 | if (crtl->tail_call_emit) |
6990 | fixup_tail_calls (); |
6991 | |
6992 | HOST_WIDE_INT patch_area_size, patch_area_entry; |
6993 | parse_and_check_patch_area (flag_patchable_function_entry, report_error: false, |
6994 | patch_area_size: &patch_area_size, patch_area_start: &patch_area_entry); |
6995 | |
6996 | tree patchable_function_entry_attr |
6997 | = lookup_attribute (attr_name: "patchable_function_entry" , |
6998 | DECL_ATTRIBUTES (cfun->decl)); |
6999 | if (patchable_function_entry_attr) |
7000 | { |
7001 | tree pp_val = TREE_VALUE (patchable_function_entry_attr); |
7002 | tree patchable_function_entry_value1 = TREE_VALUE (pp_val); |
7003 | |
7004 | patch_area_size = tree_to_uhwi (patchable_function_entry_value1); |
7005 | patch_area_entry = 0; |
7006 | if (TREE_CHAIN (pp_val) != NULL_TREE) |
7007 | { |
7008 | tree patchable_function_entry_value2 |
7009 | = TREE_VALUE (TREE_CHAIN (pp_val)); |
7010 | patch_area_entry = tree_to_uhwi (patchable_function_entry_value2); |
7011 | } |
7012 | } |
7013 | |
7014 | if (patch_area_entry > patch_area_size) |
7015 | { |
7016 | if (patch_area_size > 0) |
7017 | warning (OPT_Wattributes, |
7018 | "patchable function entry %wu exceeds size %wu" , |
7019 | patch_area_entry, patch_area_size); |
7020 | patch_area_entry = 0; |
7021 | } |
7022 | |
7023 | crtl->patch_area_size = patch_area_size; |
7024 | crtl->patch_area_entry = patch_area_entry; |
7025 | |
7026 | /* BB subdivision may have created basic blocks that are only reachable |
7027 | from unlikely bbs but not marked as such in the profile. */ |
7028 | if (optimize) |
7029 | propagate_unlikely_bbs_forward (); |
7030 | |
7031 | /* Remove unreachable blocks, otherwise we cannot compute dominators |
7032 | which are needed for loop state verification. As a side-effect |
7033 | this also compacts blocks. |
7034 | ??? We cannot remove trivially dead insns here as for example |
7035 | the DRAP reg on i?86 is not magically live at this point. |
7036 | gcc.c-torture/execute/ipa-sra-2.c execution, -Os -m32 fails otherwise. */ |
7037 | cleanup_cfg (CLEANUP_NO_INSN_DEL); |
7038 | |
7039 | checking_verify_flow_info (); |
7040 | |
7041 | /* Initialize pseudos allocated for hard registers. */ |
7042 | emit_initial_value_sets (); |
7043 | |
7044 | /* And finally unshare all RTL. */ |
7045 | unshare_all_rtl (); |
7046 | |
7047 | /* There's no need to defer outputting this function any more; we |
7048 | know we want to output it. */ |
7049 | DECL_DEFER_OUTPUT (current_function_decl) = 0; |
7050 | |
7051 | /* Now that we're done expanding trees to RTL, we shouldn't have any |
7052 | more CONCATs anywhere. */ |
7053 | generating_concat_p = 0; |
7054 | |
7055 | if (dump_file) |
7056 | { |
7057 | fprintf (stream: dump_file, |
7058 | format: "\n\n;;\n;; Full RTL generated for this function:\n;;\n" ); |
7059 | /* And the pass manager will dump RTL for us. */ |
7060 | } |
7061 | |
7062 | /* If we're emitting a nested function, make sure its parent gets |
7063 | emitted as well. Doing otherwise confuses debug info. */ |
7064 | { |
7065 | tree parent; |
7066 | for (parent = DECL_CONTEXT (current_function_decl); |
7067 | parent != NULL_TREE; |
7068 | parent = get_containing_scope (parent)) |
7069 | if (TREE_CODE (parent) == FUNCTION_DECL) |
7070 | TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (parent)) = 1; |
7071 | } |
7072 | |
7073 | TREE_ASM_WRITTEN (current_function_decl) = 1; |
7074 | |
7075 | /* After expanding, the return labels are no longer needed. */ |
7076 | return_label = NULL; |
7077 | naked_return_label = NULL; |
7078 | |
7079 | /* After expanding, the tm_restart map is no longer needed. */ |
7080 | if (fun->gimple_df->tm_restart) |
7081 | fun->gimple_df->tm_restart = NULL; |
7082 | |
7083 | /* Tag the blocks with a depth number so that change_scope can find |
7084 | the common parent easily. */ |
7085 | set_block_levels (DECL_INITIAL (fun->decl), level: 0); |
7086 | default_rtl_profile (); |
7087 | |
7088 | /* For -dx discard loops now, otherwise IL verify in clean_state will |
7089 | ICE. */ |
7090 | if (rtl_dump_and_exit) |
7091 | { |
7092 | cfun->curr_properties &= ~PROP_loops; |
7093 | loop_optimizer_finalize (); |
7094 | } |
7095 | |
7096 | timevar_pop (tv: TV_POST_EXPAND); |
7097 | |
7098 | return 0; |
7099 | } |
7100 | |
7101 | } // anon namespace |
7102 | |
7103 | rtl_opt_pass * |
7104 | make_pass_expand (gcc::context *ctxt) |
7105 | { |
7106 | return new pass_expand (ctxt); |
7107 | } |
7108 | |