1/* AddressSanitizer, a fast memory error detector.
2 Copyright (C) 2012-2024 Free Software Foundation, Inc.
3 Contributed by Kostya Serebryany <kcc@google.com>
4
5This file is part of GCC.
6
7GCC is free software; you can redistribute it and/or modify it under
8the terms of the GNU General Public License as published by the Free
9Software Foundation; either version 3, or (at your option) any later
10version.
11
12GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13WARRANTY; without even the implied warranty of MERCHANTABILITY or
14FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15for more details.
16
17You should have received a copy of the GNU General Public License
18along with GCC; see the file COPYING3. If not see
19<http://www.gnu.org/licenses/>. */
20
21
22#include "config.h"
23#include "system.h"
24#include "coretypes.h"
25#include "backend.h"
26#include "target.h"
27#include "rtl.h"
28#include "tree.h"
29#include "gimple.h"
30#include "cfghooks.h"
31#include "alloc-pool.h"
32#include "tree-pass.h"
33#include "memmodel.h"
34#include "tm_p.h"
35#include "ssa.h"
36#include "stringpool.h"
37#include "tree-ssanames.h"
38#include "optabs.h"
39#include "emit-rtl.h"
40#include "cgraph.h"
41#include "gimple-pretty-print.h"
42#include "alias.h"
43#include "fold-const.h"
44#include "cfganal.h"
45#include "gimplify.h"
46#include "gimple-iterator.h"
47#include "varasm.h"
48#include "stor-layout.h"
49#include "tree-iterator.h"
50#include "stringpool.h"
51#include "attribs.h"
52#include "asan.h"
53#include "dojump.h"
54#include "explow.h"
55#include "expr.h"
56#include "output.h"
57#include "langhooks.h"
58#include "cfgloop.h"
59#include "gimple-builder.h"
60#include "gimple-fold.h"
61#include "ubsan.h"
62#include "builtins.h"
63#include "fnmatch.h"
64#include "tree-inline.h"
65#include "tree-ssa.h"
66#include "tree-eh.h"
67#include "diagnostic-core.h"
68
69/* AddressSanitizer finds out-of-bounds and use-after-free bugs
70 with <2x slowdown on average.
71
72 The tool consists of two parts:
73 instrumentation module (this file) and a run-time library.
74 The instrumentation module adds a run-time check before every memory insn.
75 For a 8- or 16- byte load accessing address X:
76 ShadowAddr = (X >> 3) + Offset
77 ShadowValue = *(char*)ShadowAddr; // *(short*) for 16-byte access.
78 if (ShadowValue)
79 __asan_report_load8(X);
80 For a load of N bytes (N=1, 2 or 4) from address X:
81 ShadowAddr = (X >> 3) + Offset
82 ShadowValue = *(char*)ShadowAddr;
83 if (ShadowValue)
84 if ((X & 7) + N - 1 > ShadowValue)
85 __asan_report_loadN(X);
86 Stores are instrumented similarly, but using __asan_report_storeN functions.
87 A call too __asan_init_vN() is inserted to the list of module CTORs.
88 N is the version number of the AddressSanitizer API. The changes between the
89 API versions are listed in libsanitizer/asan/asan_interface_internal.h.
90
91 The run-time library redefines malloc (so that redzone are inserted around
92 the allocated memory) and free (so that reuse of free-ed memory is delayed),
93 provides __asan_report* and __asan_init_vN functions.
94
95 Read more:
96 http://code.google.com/p/address-sanitizer/wiki/AddressSanitizerAlgorithm
97
98 The current implementation supports detection of out-of-bounds and
99 use-after-free in the heap, on the stack and for global variables.
100
101 [Protection of stack variables]
102
103 To understand how detection of out-of-bounds and use-after-free works
104 for stack variables, lets look at this example on x86_64 where the
105 stack grows downward:
106
107 int
108 foo ()
109 {
110 char a[24] = {0};
111 int b[2] = {0};
112
113 a[5] = 1;
114 b[1] = 2;
115
116 return a[5] + b[1];
117 }
118
119 For this function, the stack protected by asan will be organized as
120 follows, from the top of the stack to the bottom:
121
122 Slot 1/ [red zone of 32 bytes called 'RIGHT RedZone']
123
124 Slot 2/ [8 bytes of red zone, that adds up to the space of 'a' to make
125 the next slot be 32 bytes aligned; this one is called Partial
126 Redzone; this 32 bytes alignment is an asan constraint]
127
128 Slot 3/ [24 bytes for variable 'a']
129
130 Slot 4/ [red zone of 32 bytes called 'Middle RedZone']
131
132 Slot 5/ [24 bytes of Partial Red Zone (similar to slot 2]
133
134 Slot 6/ [8 bytes for variable 'b']
135
136 Slot 7/ [32 bytes of Red Zone at the bottom of the stack, called
137 'LEFT RedZone']
138
139 The 32 bytes of LEFT red zone at the bottom of the stack can be
140 decomposed as such:
141
142 1/ The first 8 bytes contain a magical asan number that is always
143 0x41B58AB3.
144
145 2/ The following 8 bytes contains a pointer to a string (to be
146 parsed at runtime by the runtime asan library), which format is
147 the following:
148
149 "<function-name> <space> <num-of-variables-on-the-stack>
150 (<32-bytes-aligned-offset-in-bytes-of-variable> <space>
151 <length-of-var-in-bytes> ){n} "
152
153 where '(...){n}' means the content inside the parenthesis occurs 'n'
154 times, with 'n' being the number of variables on the stack.
155
156 3/ The following 8 bytes contain the PC of the current function which
157 will be used by the run-time library to print an error message.
158
159 4/ The following 8 bytes are reserved for internal use by the run-time.
160
161 The shadow memory for that stack layout is going to look like this:
162
163 - content of shadow memory 8 bytes for slot 7: 0xF1F1F1F1.
164 The F1 byte pattern is a magic number called
165 ASAN_STACK_MAGIC_LEFT and is a way for the runtime to know that
166 the memory for that shadow byte is part of a the LEFT red zone
167 intended to seat at the bottom of the variables on the stack.
168
169 - content of shadow memory 8 bytes for slots 6 and 5:
170 0xF4F4F400. The F4 byte pattern is a magic number
171 called ASAN_STACK_MAGIC_PARTIAL. It flags the fact that the
172 memory region for this shadow byte is a PARTIAL red zone
173 intended to pad a variable A, so that the slot following
174 {A,padding} is 32 bytes aligned.
175
176 Note that the fact that the least significant byte of this
177 shadow memory content is 00 means that 8 bytes of its
178 corresponding memory (which corresponds to the memory of
179 variable 'b') is addressable.
180
181 - content of shadow memory 8 bytes for slot 4: 0xF2F2F2F2.
182 The F2 byte pattern is a magic number called
183 ASAN_STACK_MAGIC_MIDDLE. It flags the fact that the memory
184 region for this shadow byte is a MIDDLE red zone intended to
185 seat between two 32 aligned slots of {variable,padding}.
186
187 - content of shadow memory 8 bytes for slot 3 and 2:
188 0xF4000000. This represents is the concatenation of
189 variable 'a' and the partial red zone following it, like what we
190 had for variable 'b'. The least significant 3 bytes being 00
191 means that the 3 bytes of variable 'a' are addressable.
192
193 - content of shadow memory 8 bytes for slot 1: 0xF3F3F3F3.
194 The F3 byte pattern is a magic number called
195 ASAN_STACK_MAGIC_RIGHT. It flags the fact that the memory
196 region for this shadow byte is a RIGHT red zone intended to seat
197 at the top of the variables of the stack.
198
199 Note that the real variable layout is done in expand_used_vars in
200 cfgexpand.cc. As far as Address Sanitizer is concerned, it lays out
201 stack variables as well as the different red zones, emits some
202 prologue code to populate the shadow memory as to poison (mark as
203 non-accessible) the regions of the red zones and mark the regions of
204 stack variables as accessible, and emit some epilogue code to
205 un-poison (mark as accessible) the regions of red zones right before
206 the function exits.
207
208 [Protection of global variables]
209
210 The basic idea is to insert a red zone between two global variables
211 and install a constructor function that calls the asan runtime to do
212 the populating of the relevant shadow memory regions at load time.
213
214 So the global variables are laid out as to insert a red zone between
215 them. The size of the red zones is so that each variable starts on a
216 32 bytes boundary.
217
218 Then a constructor function is installed so that, for each global
219 variable, it calls the runtime asan library function
220 __asan_register_globals_with an instance of this type:
221
222 struct __asan_global
223 {
224 // Address of the beginning of the global variable.
225 const void *__beg;
226
227 // Initial size of the global variable.
228 uptr __size;
229
230 // Size of the global variable + size of the red zone. This
231 // size is 32 bytes aligned.
232 uptr __size_with_redzone;
233
234 // Name of the global variable.
235 const void *__name;
236
237 // Name of the module where the global variable is declared.
238 const void *__module_name;
239
240 // 1 if it has dynamic initialization, 0 otherwise.
241 uptr __has_dynamic_init;
242
243 // A pointer to struct that contains source location, could be NULL.
244 __asan_global_source_location *__location;
245 }
246
247 A destructor function that calls the runtime asan library function
248 _asan_unregister_globals is also installed. */
249
250static unsigned HOST_WIDE_INT asan_shadow_offset_value;
251static bool asan_shadow_offset_computed;
252static vec<char *> sanitized_sections;
253static tree last_alloca_addr;
254
255/* Set of variable declarations that are going to be guarded by
256 use-after-scope sanitizer. */
257
258hash_set<tree> *asan_handled_variables = NULL;
259
260hash_set <tree> *asan_used_labels = NULL;
261
262/* Global variables for HWASAN stack tagging. */
263/* hwasan_frame_tag_offset records the offset from the frame base tag that the
264 next object should have. */
265static uint8_t hwasan_frame_tag_offset = 0;
266/* hwasan_frame_base_ptr is a pointer with the same address as
267 `virtual_stack_vars_rtx` for the current frame, and with the frame base tag
268 stored in it. N.b. this global RTX does not need to be marked GTY, but is
269 done so anyway. The need is not there since all uses are in just one pass
270 (cfgexpand) and there are no calls to ggc_collect between the uses. We mark
271 it GTY(()) anyway to allow the use of the variable later on if needed by
272 future features. */
273static GTY(()) rtx hwasan_frame_base_ptr = NULL_RTX;
274/* hwasan_frame_base_init_seq is the sequence of RTL insns that will initialize
275 the hwasan_frame_base_ptr. When the hwasan_frame_base_ptr is requested, we
276 generate this sequence but do not emit it. If the sequence was created it
277 is emitted once the function body has been expanded.
278
279 This delay is because the frame base pointer may be needed anywhere in the
280 function body, or needed by the expand_used_vars function. Emitting once in
281 a known place is simpler than requiring the emission of the instructions to
282 be know where it should go depending on the first place the hwasan frame
283 base is needed. */
284static GTY(()) rtx_insn *hwasan_frame_base_init_seq = NULL;
285
286/* Structure defining the extent of one object on the stack that HWASAN needs
287 to tag in the corresponding shadow stack space.
288
289 The range this object spans on the stack is between `untagged_base +
290 nearest_offset` and `untagged_base + farthest_offset`.
291 `tagged_base` is an rtx containing the same value as `untagged_base` but
292 with a random tag stored in the top byte. We record both `untagged_base`
293 and `tagged_base` so that `hwasan_emit_prologue` can use both without having
294 to emit RTL into the instruction stream to re-calculate one from the other.
295 (`hwasan_emit_prologue` needs to use both bases since the
296 __hwasan_tag_memory call it emits uses an untagged value, and it calculates
297 the tag to store in shadow memory based on the tag_offset plus the tag in
298 tagged_base). */
299struct hwasan_stack_var
300{
301 rtx untagged_base;
302 rtx tagged_base;
303 poly_int64 nearest_offset;
304 poly_int64 farthest_offset;
305 uint8_t tag_offset;
306};
307
308/* Variable recording all stack variables that HWASAN needs to tag.
309 Does not need to be marked as GTY(()) since every use is in the cfgexpand
310 pass and gcc_collect is not called in the middle of that pass. */
311static vec<hwasan_stack_var> hwasan_tagged_stack_vars;
312
313
314/* Sets shadow offset to value in string VAL. */
315
316bool
317set_asan_shadow_offset (const char *val)
318{
319 char *endp;
320
321 errno = 0;
322#ifdef HAVE_LONG_LONG
323 asan_shadow_offset_value = strtoull (nptr: val, endptr: &endp, base: 0);
324#else
325 asan_shadow_offset_value = strtoul (val, &endp, 0);
326#endif
327 if (!(*val != '\0' && *endp == '\0' && errno == 0))
328 return false;
329
330 asan_shadow_offset_computed = true;
331
332 return true;
333}
334
335/* Set list of user-defined sections that need to be sanitized. */
336
337void
338set_sanitized_sections (const char *sections)
339{
340 char *pat;
341 unsigned i;
342 FOR_EACH_VEC_ELT (sanitized_sections, i, pat)
343 free (ptr: pat);
344 sanitized_sections.truncate (size: 0);
345
346 for (const char *s = sections; *s; )
347 {
348 const char *end;
349 for (end = s; *end && *end != ','; ++end);
350 size_t len = end - s;
351 sanitized_sections.safe_push (obj: xstrndup (s, len));
352 s = *end ? end + 1 : end;
353 }
354}
355
356bool
357asan_mark_p (gimple *stmt, enum asan_mark_flags flag)
358{
359 return (gimple_call_internal_p (gs: stmt, fn: IFN_ASAN_MARK)
360 && tree_to_uhwi (gimple_call_arg (gs: stmt, index: 0)) == flag);
361}
362
363bool
364asan_sanitize_stack_p (void)
365{
366 return (sanitize_flags_p (flag: SANITIZE_ADDRESS) && param_asan_stack);
367}
368
369bool
370asan_sanitize_allocas_p (void)
371{
372 return (asan_sanitize_stack_p () && param_asan_protect_allocas);
373}
374
375bool
376asan_instrument_reads (void)
377{
378 return (sanitize_flags_p (flag: SANITIZE_ADDRESS) && param_asan_instrument_reads);
379}
380
381bool
382asan_instrument_writes (void)
383{
384 return (sanitize_flags_p (flag: SANITIZE_ADDRESS) && param_asan_instrument_writes);
385}
386
387bool
388asan_memintrin (void)
389{
390 return (sanitize_flags_p (flag: SANITIZE_ADDRESS) && param_asan_memintrin);
391}
392
393
394/* Support for --param asan-kernel-mem-intrinsic-prefix=1. */
395static GTY(()) rtx asan_memfn_rtls[3];
396
397rtx
398asan_memfn_rtl (tree fndecl)
399{
400 int i;
401 const char *f, *p;
402 char buf[sizeof ("__hwasan_memmove")];
403
404 switch (DECL_FUNCTION_CODE (decl: fndecl))
405 {
406 case BUILT_IN_MEMCPY: i = 0; f = "memcpy"; break;
407 case BUILT_IN_MEMSET: i = 1; f = "memset"; break;
408 case BUILT_IN_MEMMOVE: i = 2; f = "memmove"; break;
409 default: gcc_unreachable ();
410 }
411 if (asan_memfn_rtls[i] == NULL_RTX)
412 {
413 tree save_name = DECL_NAME (fndecl);
414 tree save_assembler_name = DECL_ASSEMBLER_NAME (fndecl);
415 rtx save_rtl = DECL_RTL (fndecl);
416 if (flag_sanitize & SANITIZE_KERNEL_HWADDRESS)
417 p = "__hwasan_";
418 else
419 p = "__asan_";
420 strcpy (dest: buf, src: p);
421 strcat (dest: buf, src: f);
422 DECL_NAME (fndecl) = get_identifier (buf);
423 DECL_ASSEMBLER_NAME_RAW (fndecl) = NULL_TREE;
424 SET_DECL_RTL (fndecl, NULL_RTX);
425 asan_memfn_rtls[i] = DECL_RTL (fndecl);
426 DECL_NAME (fndecl) = save_name;
427 DECL_ASSEMBLER_NAME_RAW (fndecl) = save_assembler_name;
428 SET_DECL_RTL (fndecl, save_rtl);
429 }
430 return asan_memfn_rtls[i];
431}
432
433
434/* Checks whether section SEC should be sanitized. */
435
436static bool
437section_sanitized_p (const char *sec)
438{
439 char *pat;
440 unsigned i;
441 FOR_EACH_VEC_ELT (sanitized_sections, i, pat)
442 if (fnmatch (pattern: pat, string: sec, FNM_PERIOD) == 0)
443 return true;
444 return false;
445}
446
447/* Returns Asan shadow offset. */
448
449static unsigned HOST_WIDE_INT
450asan_shadow_offset ()
451{
452 if (!asan_shadow_offset_computed)
453 {
454 asan_shadow_offset_computed = true;
455 asan_shadow_offset_value = targetm.asan_shadow_offset ();
456 }
457 return asan_shadow_offset_value;
458}
459
460/* Returns Asan shadow offset has been set. */
461bool
462asan_shadow_offset_set_p ()
463{
464 return asan_shadow_offset_computed;
465}
466
467alias_set_type asan_shadow_set = -1;
468
469/* Pointer types to 1, 2 or 4 byte integers in shadow memory. A separate
470 alias set is used for all shadow memory accesses. */
471static GTY(()) tree shadow_ptr_types[3];
472
473/* Decl for __asan_option_detect_stack_use_after_return. */
474static GTY(()) tree asan_detect_stack_use_after_return;
475
476/* Hashtable support for memory references used by gimple
477 statements. */
478
479/* This type represents a reference to a memory region. */
480struct asan_mem_ref
481{
482 /* The expression of the beginning of the memory region. */
483 tree start;
484
485 /* The size of the access. */
486 HOST_WIDE_INT access_size;
487};
488
489object_allocator <asan_mem_ref> asan_mem_ref_pool ("asan_mem_ref");
490
491/* Initializes an instance of asan_mem_ref. */
492
493static void
494asan_mem_ref_init (asan_mem_ref *ref, tree start, HOST_WIDE_INT access_size)
495{
496 ref->start = start;
497 ref->access_size = access_size;
498}
499
500/* Allocates memory for an instance of asan_mem_ref into the memory
501 pool returned by asan_mem_ref_get_alloc_pool and initialize it.
502 START is the address of (or the expression pointing to) the
503 beginning of memory reference. ACCESS_SIZE is the size of the
504 access to the referenced memory. */
505
506static asan_mem_ref*
507asan_mem_ref_new (tree start, HOST_WIDE_INT access_size)
508{
509 asan_mem_ref *ref = asan_mem_ref_pool.allocate ();
510
511 asan_mem_ref_init (ref, start, access_size);
512 return ref;
513}
514
515/* This builds and returns a pointer to the end of the memory region
516 that starts at START and of length LEN. */
517
518tree
519asan_mem_ref_get_end (tree start, tree len)
520{
521 if (len == NULL_TREE || integer_zerop (len))
522 return start;
523
524 if (!ptrofftype_p (type: len))
525 len = convert_to_ptrofftype (len);
526
527 return fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (start), start, len);
528}
529
530/* Return a tree expression that represents the end of the referenced
531 memory region. Beware that this function can actually build a new
532 tree expression. */
533
534tree
535asan_mem_ref_get_end (const asan_mem_ref *ref, tree len)
536{
537 return asan_mem_ref_get_end (start: ref->start, len);
538}
539
540struct asan_mem_ref_hasher : nofree_ptr_hash <asan_mem_ref>
541{
542 static inline hashval_t hash (const asan_mem_ref *);
543 static inline bool equal (const asan_mem_ref *, const asan_mem_ref *);
544};
545
546/* Hash a memory reference. */
547
548inline hashval_t
549asan_mem_ref_hasher::hash (const asan_mem_ref *mem_ref)
550{
551 return iterative_hash_expr (tree: mem_ref->start, seed: 0);
552}
553
554/* Compare two memory references. We accept the length of either
555 memory references to be NULL_TREE. */
556
557inline bool
558asan_mem_ref_hasher::equal (const asan_mem_ref *m1,
559 const asan_mem_ref *m2)
560{
561 return operand_equal_p (m1->start, m2->start, flags: 0);
562}
563
564static hash_table<asan_mem_ref_hasher> *asan_mem_ref_ht;
565
566/* Returns a reference to the hash table containing memory references.
567 This function ensures that the hash table is created. Note that
568 this hash table is updated by the function
569 update_mem_ref_hash_table. */
570
571static hash_table<asan_mem_ref_hasher> *
572get_mem_ref_hash_table ()
573{
574 if (!asan_mem_ref_ht)
575 asan_mem_ref_ht = new hash_table<asan_mem_ref_hasher> (10);
576
577 return asan_mem_ref_ht;
578}
579
580/* Clear all entries from the memory references hash table. */
581
582static void
583empty_mem_ref_hash_table ()
584{
585 if (asan_mem_ref_ht)
586 asan_mem_ref_ht->empty ();
587}
588
589/* Free the memory references hash table. */
590
591static void
592free_mem_ref_resources ()
593{
594 delete asan_mem_ref_ht;
595 asan_mem_ref_ht = NULL;
596
597 asan_mem_ref_pool.release ();
598}
599
600/* Return true iff the memory reference REF has been instrumented. */
601
602static bool
603has_mem_ref_been_instrumented (tree ref, HOST_WIDE_INT access_size)
604{
605 asan_mem_ref r;
606 asan_mem_ref_init (ref: &r, start: ref, access_size);
607
608 asan_mem_ref *saved_ref = get_mem_ref_hash_table ()->find (value: &r);
609 return saved_ref && saved_ref->access_size >= access_size;
610}
611
612/* Return true iff the memory reference REF has been instrumented. */
613
614static bool
615has_mem_ref_been_instrumented (const asan_mem_ref *ref)
616{
617 return has_mem_ref_been_instrumented (ref: ref->start, access_size: ref->access_size);
618}
619
620/* Return true iff access to memory region starting at REF and of
621 length LEN has been instrumented. */
622
623static bool
624has_mem_ref_been_instrumented (const asan_mem_ref *ref, tree len)
625{
626 HOST_WIDE_INT size_in_bytes
627 = tree_fits_shwi_p (len) ? tree_to_shwi (len) : -1;
628
629 return size_in_bytes != -1
630 && has_mem_ref_been_instrumented (ref: ref->start, access_size: size_in_bytes);
631}
632
633/* Set REF to the memory reference present in a gimple assignment
634 ASSIGNMENT. Return true upon successful completion, false
635 otherwise. */
636
637static bool
638get_mem_ref_of_assignment (const gassign *assignment,
639 asan_mem_ref *ref,
640 bool *ref_is_store)
641{
642 gcc_assert (gimple_assign_single_p (assignment));
643
644 if (gimple_store_p (gs: assignment)
645 && !gimple_clobber_p (s: assignment))
646 {
647 ref->start = gimple_assign_lhs (gs: assignment);
648 *ref_is_store = true;
649 }
650 else if (gimple_assign_load_p (assignment))
651 {
652 ref->start = gimple_assign_rhs1 (gs: assignment);
653 *ref_is_store = false;
654 }
655 else
656 return false;
657
658 ref->access_size = int_size_in_bytes (TREE_TYPE (ref->start));
659 return true;
660}
661
662/* Return address of last allocated dynamic alloca. */
663
664static tree
665get_last_alloca_addr ()
666{
667 if (last_alloca_addr)
668 return last_alloca_addr;
669
670 last_alloca_addr = create_tmp_reg (ptr_type_node, "last_alloca_addr");
671 gassign *g = gimple_build_assign (last_alloca_addr, null_pointer_node);
672 edge e = single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun));
673 gsi_insert_on_edge_immediate (e, g);
674 return last_alloca_addr;
675}
676
677/* Insert __asan_allocas_unpoison (top, bottom) call before
678 __builtin_stack_restore (new_sp) call.
679 The pseudocode of this routine should look like this:
680 top = last_alloca_addr;
681 bot = new_sp;
682 __asan_allocas_unpoison (top, bot);
683 last_alloca_addr = new_sp;
684 __builtin_stack_restore (new_sp);
685 In general, we can't use new_sp as bot parameter because on some
686 architectures SP has non zero offset from dynamic stack area. Moreover, on
687 some architectures this offset (STACK_DYNAMIC_OFFSET) becomes known for each
688 particular function only after all callees were expanded to rtl.
689 The most noticeable example is PowerPC{,64}, see
690 http://refspecs.linuxfoundation.org/ELF/ppc64/PPC-elf64abi.html#DYNAM-STACK.
691 To overcome the issue we use following trick: pass new_sp as a second
692 parameter to __asan_allocas_unpoison and rewrite it during expansion with
693 new_sp + (virtual_dynamic_stack_rtx - sp) later in
694 expand_asan_emit_allocas_unpoison function.
695
696 HWASAN needs to do very similar, the eventual pseudocode should be:
697 __hwasan_tag_memory (virtual_stack_dynamic_rtx,
698 0,
699 new_sp - sp);
700 __builtin_stack_restore (new_sp)
701
702 Need to use the same trick to handle STACK_DYNAMIC_OFFSET as described
703 above. */
704
705static void
706handle_builtin_stack_restore (gcall *call, gimple_stmt_iterator *iter)
707{
708 if (!iter
709 || !(asan_sanitize_allocas_p () || hwasan_sanitize_allocas_p ()))
710 return;
711
712 tree restored_stack = gimple_call_arg (gs: call, index: 0);
713
714 gimple *g;
715
716 if (hwasan_sanitize_allocas_p ())
717 {
718 enum internal_fn fn = IFN_HWASAN_ALLOCA_UNPOISON;
719 /* There is only one piece of information `expand_HWASAN_ALLOCA_UNPOISON`
720 needs to work. This is the length of the area that we're
721 deallocating. Since the stack pointer is known at expand time, the
722 position of the new stack pointer after deallocation is enough
723 information to calculate this length. */
724 g = gimple_build_call_internal (fn, 1, restored_stack);
725 }
726 else
727 {
728 tree last_alloca = get_last_alloca_addr ();
729 tree fn = builtin_decl_implicit (fncode: BUILT_IN_ASAN_ALLOCAS_UNPOISON);
730 g = gimple_build_call (fn, 2, last_alloca, restored_stack);
731 gsi_insert_before (iter, g, GSI_SAME_STMT);
732 g = gimple_build_assign (last_alloca, restored_stack);
733 }
734
735 gsi_insert_before (iter, g, GSI_SAME_STMT);
736}
737
738/* Deploy and poison redzones around __builtin_alloca call. To do this, we
739 should replace this call with another one with changed parameters and
740 replace all its uses with new address, so
741 addr = __builtin_alloca (old_size, align);
742 is replaced by
743 left_redzone_size = max (align, ASAN_RED_ZONE_SIZE);
744 Following two statements are optimized out if we know that
745 old_size & (ASAN_RED_ZONE_SIZE - 1) == 0, i.e. alloca doesn't need partial
746 redzone.
747 misalign = old_size & (ASAN_RED_ZONE_SIZE - 1);
748 partial_redzone_size = ASAN_RED_ZONE_SIZE - misalign;
749 right_redzone_size = ASAN_RED_ZONE_SIZE;
750 additional_size = left_redzone_size + partial_redzone_size +
751 right_redzone_size;
752 new_size = old_size + additional_size;
753 new_alloca = __builtin_alloca (new_size, max (align, 32))
754 __asan_alloca_poison (new_alloca, old_size)
755 addr = new_alloca + max (align, ASAN_RED_ZONE_SIZE);
756 last_alloca_addr = new_alloca;
757 ADDITIONAL_SIZE is added to make new memory allocation contain not only
758 requested memory, but also left, partial and right redzones as well as some
759 additional space, required by alignment. */
760
761static void
762handle_builtin_alloca (gcall *call, gimple_stmt_iterator *iter)
763{
764 if (!iter
765 || !(asan_sanitize_allocas_p () || hwasan_sanitize_allocas_p ()))
766 return;
767
768 gassign *g;
769 gcall *gg;
770 tree callee = gimple_call_fndecl (gs: call);
771 tree lhs = gimple_call_lhs (gs: call);
772 tree old_size = gimple_call_arg (gs: call, index: 0);
773 tree ptr_type = lhs ? TREE_TYPE (lhs) : ptr_type_node;
774 tree partial_size = NULL_TREE;
775 unsigned int align
776 = DECL_FUNCTION_CODE (decl: callee) == BUILT_IN_ALLOCA
777 ? 0 : tree_to_uhwi (gimple_call_arg (gs: call, index: 1));
778
779 bool throws = false;
780 edge e = NULL;
781 if (stmt_can_throw_internal (cfun, call))
782 {
783 if (!lhs)
784 return;
785 throws = true;
786 e = find_fallthru_edge (edges: gsi_bb (i: *iter)->succs);
787 }
788
789 if (hwasan_sanitize_allocas_p ())
790 {
791 gimple_seq stmts = NULL;
792 location_t loc = gimple_location (g: gsi_stmt (i: *iter));
793 /*
794 HWASAN needs a different expansion.
795
796 addr = __builtin_alloca (size, align);
797
798 should be replaced by
799
800 new_size = size rounded up to HWASAN_TAG_GRANULE_SIZE byte alignment;
801 untagged_addr = __builtin_alloca (new_size, align);
802 tag = __hwasan_choose_alloca_tag ();
803 addr = ifn_HWASAN_SET_TAG (untagged_addr, tag);
804 __hwasan_tag_memory (untagged_addr, tag, new_size);
805 */
806 /* Ensure alignment at least HWASAN_TAG_GRANULE_SIZE bytes so we start on
807 a tag granule. */
808 align = align > HWASAN_TAG_GRANULE_SIZE ? align : HWASAN_TAG_GRANULE_SIZE;
809
810 tree old_size = gimple_call_arg (gs: call, index: 0);
811 tree new_size = gimple_build_round_up (seq: &stmts, loc, size_type_node,
812 old_size,
813 HWASAN_TAG_GRANULE_SIZE);
814
815 /* Make the alloca call */
816 tree untagged_addr
817 = gimple_build (seq: &stmts, loc,
818 fn: as_combined_fn (fn: BUILT_IN_ALLOCA_WITH_ALIGN), type: ptr_type,
819 args: new_size, args: build_int_cst (size_type_node, align));
820
821 /* Choose the tag.
822 Here we use an internal function so we can choose the tag at expand
823 time. We need the decision to be made after stack variables have been
824 assigned their tag (i.e. once the hwasan_frame_tag_offset variable has
825 been set to one after the last stack variables tag). */
826 tree tag = gimple_build (seq: &stmts, loc, fn: CFN_HWASAN_CHOOSE_TAG,
827 unsigned_char_type_node);
828
829 /* Add tag to pointer. */
830 tree addr
831 = gimple_build (seq: &stmts, loc, fn: CFN_HWASAN_SET_TAG, type: ptr_type,
832 args: untagged_addr, args: tag);
833
834 /* Tag shadow memory.
835 NOTE: require using `untagged_addr` here for libhwasan API. */
836 gimple_build (seq: &stmts, loc, fn: as_combined_fn (fn: BUILT_IN_HWASAN_TAG_MEM),
837 void_type_node, args: untagged_addr, args: tag, args: new_size);
838
839 /* Insert the built up code sequence into the original instruction stream
840 the iterator points to. */
841 gsi_insert_seq_before (iter, stmts, GSI_SAME_STMT);
842
843 /* Finally, replace old alloca ptr with NEW_ALLOCA. */
844 replace_call_with_value (iter, addr);
845 return;
846 }
847
848 tree last_alloca = get_last_alloca_addr ();
849 const HOST_WIDE_INT redzone_mask = ASAN_RED_ZONE_SIZE - 1;
850
851 /* If ALIGN > ASAN_RED_ZONE_SIZE, we embed left redzone into first ALIGN
852 bytes of allocated space. Otherwise, align alloca to ASAN_RED_ZONE_SIZE
853 manually. */
854 align = MAX (align, ASAN_RED_ZONE_SIZE * BITS_PER_UNIT);
855
856 tree alloca_rz_mask = build_int_cst (size_type_node, redzone_mask);
857 tree redzone_size = build_int_cst (size_type_node, ASAN_RED_ZONE_SIZE);
858
859 /* Extract lower bits from old_size. */
860 wide_int size_nonzero_bits = get_nonzero_bits (old_size);
861 wide_int rz_mask
862 = wi::uhwi (val: redzone_mask, precision: wi::get_precision (x: size_nonzero_bits));
863 wide_int old_size_lower_bits = wi::bit_and (x: size_nonzero_bits, y: rz_mask);
864
865 /* If alloca size is aligned to ASAN_RED_ZONE_SIZE, we don't need partial
866 redzone. Otherwise, compute its size here. */
867 if (wi::ne_p (x: old_size_lower_bits, y: 0))
868 {
869 /* misalign = size & (ASAN_RED_ZONE_SIZE - 1)
870 partial_size = ASAN_RED_ZONE_SIZE - misalign. */
871 g = gimple_build_assign (make_ssa_name (size_type_node, NULL),
872 BIT_AND_EXPR, old_size, alloca_rz_mask);
873 gsi_insert_before (iter, g, GSI_SAME_STMT);
874 tree misalign = gimple_assign_lhs (gs: g);
875 g = gimple_build_assign (make_ssa_name (size_type_node, NULL), MINUS_EXPR,
876 redzone_size, misalign);
877 gsi_insert_before (iter, g, GSI_SAME_STMT);
878 partial_size = gimple_assign_lhs (gs: g);
879 }
880
881 /* additional_size = align + ASAN_RED_ZONE_SIZE. */
882 tree additional_size = build_int_cst (size_type_node, align / BITS_PER_UNIT
883 + ASAN_RED_ZONE_SIZE);
884 /* If alloca has partial redzone, include it to additional_size too. */
885 if (partial_size)
886 {
887 /* additional_size += partial_size. */
888 g = gimple_build_assign (make_ssa_name (size_type_node), PLUS_EXPR,
889 partial_size, additional_size);
890 gsi_insert_before (iter, g, GSI_SAME_STMT);
891 additional_size = gimple_assign_lhs (gs: g);
892 }
893
894 /* new_size = old_size + additional_size. */
895 g = gimple_build_assign (make_ssa_name (size_type_node), PLUS_EXPR, old_size,
896 additional_size);
897 gsi_insert_before (iter, g, GSI_SAME_STMT);
898 tree new_size = gimple_assign_lhs (gs: g);
899
900 /* Build new __builtin_alloca call:
901 new_alloca_with_rz = __builtin_alloca (new_size, align). */
902 tree fn = builtin_decl_implicit (fncode: BUILT_IN_ALLOCA_WITH_ALIGN);
903 gg = gimple_build_call (fn, 2, new_size,
904 build_int_cst (size_type_node, align));
905 tree new_alloca_with_rz = make_ssa_name (var: ptr_type, stmt: gg);
906 gimple_call_set_lhs (gs: gg, lhs: new_alloca_with_rz);
907 if (throws)
908 {
909 gimple_call_set_lhs (gs: call, NULL);
910 gsi_replace (iter, gg, true);
911 }
912 else
913 gsi_insert_before (iter, gg, GSI_SAME_STMT);
914
915 /* new_alloca = new_alloca_with_rz + align. */
916 g = gimple_build_assign (make_ssa_name (var: ptr_type), POINTER_PLUS_EXPR,
917 new_alloca_with_rz,
918 build_int_cst (size_type_node,
919 align / BITS_PER_UNIT));
920 gimple_stmt_iterator gsi = gsi_none ();
921 if (throws)
922 {
923 gsi_insert_on_edge_immediate (e, g);
924 gsi = gsi_for_stmt (g);
925 }
926 else
927 gsi_insert_before (iter, g, GSI_SAME_STMT);
928 tree new_alloca = gimple_assign_lhs (gs: g);
929
930 /* Poison newly created alloca redzones:
931 __asan_alloca_poison (new_alloca, old_size). */
932 fn = builtin_decl_implicit (fncode: BUILT_IN_ASAN_ALLOCA_POISON);
933 gg = gimple_build_call (fn, 2, new_alloca, old_size);
934 if (throws)
935 gsi_insert_after (&gsi, gg, GSI_NEW_STMT);
936 else
937 gsi_insert_before (iter, gg, GSI_SAME_STMT);
938
939 /* Save new_alloca_with_rz value into last_alloca to use it during
940 allocas unpoisoning. */
941 g = gimple_build_assign (last_alloca, new_alloca_with_rz);
942 if (throws)
943 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
944 else
945 gsi_insert_before (iter, g, GSI_SAME_STMT);
946
947 /* Finally, replace old alloca ptr with NEW_ALLOCA. */
948 if (throws)
949 {
950 g = gimple_build_assign (lhs, new_alloca);
951 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
952 }
953 else
954 replace_call_with_value (iter, new_alloca);
955}
956
957/* Return the memory references contained in a gimple statement
958 representing a builtin call that has to do with memory access. */
959
960static bool
961get_mem_refs_of_builtin_call (gcall *call,
962 asan_mem_ref *src0,
963 tree *src0_len,
964 bool *src0_is_store,
965 asan_mem_ref *src1,
966 tree *src1_len,
967 bool *src1_is_store,
968 asan_mem_ref *dst,
969 tree *dst_len,
970 bool *dst_is_store,
971 bool *dest_is_deref,
972 bool *intercepted_p,
973 gimple_stmt_iterator *iter = NULL)
974{
975 gcc_checking_assert (gimple_call_builtin_p (call, BUILT_IN_NORMAL));
976
977 tree callee = gimple_call_fndecl (gs: call);
978 tree source0 = NULL_TREE, source1 = NULL_TREE,
979 dest = NULL_TREE, len = NULL_TREE;
980 bool is_store = true, got_reference_p = false;
981 HOST_WIDE_INT access_size = 1;
982
983 *intercepted_p = asan_intercepted_p (fcode: (DECL_FUNCTION_CODE (decl: callee)));
984
985 switch (DECL_FUNCTION_CODE (decl: callee))
986 {
987 /* (s, s, n) style memops. */
988 case BUILT_IN_BCMP:
989 case BUILT_IN_MEMCMP:
990 source0 = gimple_call_arg (gs: call, index: 0);
991 source1 = gimple_call_arg (gs: call, index: 1);
992 len = gimple_call_arg (gs: call, index: 2);
993 break;
994
995 /* (src, dest, n) style memops. */
996 case BUILT_IN_BCOPY:
997 source0 = gimple_call_arg (gs: call, index: 0);
998 dest = gimple_call_arg (gs: call, index: 1);
999 len = gimple_call_arg (gs: call, index: 2);
1000 break;
1001
1002 /* (dest, src, n) style memops. */
1003 case BUILT_IN_MEMCPY:
1004 case BUILT_IN_MEMCPY_CHK:
1005 case BUILT_IN_MEMMOVE:
1006 case BUILT_IN_MEMMOVE_CHK:
1007 case BUILT_IN_MEMPCPY:
1008 case BUILT_IN_MEMPCPY_CHK:
1009 dest = gimple_call_arg (gs: call, index: 0);
1010 source0 = gimple_call_arg (gs: call, index: 1);
1011 len = gimple_call_arg (gs: call, index: 2);
1012 break;
1013
1014 /* (dest, n) style memops. */
1015 case BUILT_IN_BZERO:
1016 dest = gimple_call_arg (gs: call, index: 0);
1017 len = gimple_call_arg (gs: call, index: 1);
1018 break;
1019
1020 /* (dest, x, n) style memops*/
1021 case BUILT_IN_MEMSET:
1022 case BUILT_IN_MEMSET_CHK:
1023 dest = gimple_call_arg (gs: call, index: 0);
1024 len = gimple_call_arg (gs: call, index: 2);
1025 break;
1026
1027 case BUILT_IN_STRLEN:
1028 /* Special case strlen here since its length is taken from its return
1029 value.
1030
1031 The approach taken by the sanitizers is to check a memory access
1032 before it's taken. For ASAN strlen is intercepted by libasan, so no
1033 check is inserted by the compiler.
1034
1035 This function still returns `true` and provides a length to the rest
1036 of the ASAN pass in order to record what areas have been checked,
1037 avoiding superfluous checks later on.
1038
1039 HWASAN does not intercept any of these internal functions.
1040 This means that checks for memory accesses must be inserted by the
1041 compiler.
1042 strlen is a special case, because we can tell the length from the
1043 return of the function, but that is not known until after the function
1044 has returned.
1045
1046 Hence we can't check the memory access before it happens.
1047 We could check the memory access after it has already happened, but
1048 for now we choose to just ignore `strlen` calls.
1049 This decision was simply made because that means the special case is
1050 limited to this one case of this one function. */
1051 if (hwasan_sanitize_p ())
1052 return false;
1053 source0 = gimple_call_arg (gs: call, index: 0);
1054 len = gimple_call_lhs (gs: call);
1055 break;
1056
1057 case BUILT_IN_STACK_RESTORE:
1058 handle_builtin_stack_restore (call, iter);
1059 break;
1060
1061 CASE_BUILT_IN_ALLOCA:
1062 handle_builtin_alloca (call, iter);
1063 break;
1064 /* And now the __atomic* and __sync builtins.
1065 These are handled differently from the classical memory
1066 access builtins above. */
1067
1068 case BUILT_IN_ATOMIC_LOAD_1:
1069 is_store = false;
1070 /* FALLTHRU */
1071 case BUILT_IN_SYNC_FETCH_AND_ADD_1:
1072 case BUILT_IN_SYNC_FETCH_AND_SUB_1:
1073 case BUILT_IN_SYNC_FETCH_AND_OR_1:
1074 case BUILT_IN_SYNC_FETCH_AND_AND_1:
1075 case BUILT_IN_SYNC_FETCH_AND_XOR_1:
1076 case BUILT_IN_SYNC_FETCH_AND_NAND_1:
1077 case BUILT_IN_SYNC_ADD_AND_FETCH_1:
1078 case BUILT_IN_SYNC_SUB_AND_FETCH_1:
1079 case BUILT_IN_SYNC_OR_AND_FETCH_1:
1080 case BUILT_IN_SYNC_AND_AND_FETCH_1:
1081 case BUILT_IN_SYNC_XOR_AND_FETCH_1:
1082 case BUILT_IN_SYNC_NAND_AND_FETCH_1:
1083 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_1:
1084 case BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_1:
1085 case BUILT_IN_SYNC_LOCK_TEST_AND_SET_1:
1086 case BUILT_IN_SYNC_LOCK_RELEASE_1:
1087 case BUILT_IN_ATOMIC_EXCHANGE_1:
1088 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_1:
1089 case BUILT_IN_ATOMIC_STORE_1:
1090 case BUILT_IN_ATOMIC_ADD_FETCH_1:
1091 case BUILT_IN_ATOMIC_SUB_FETCH_1:
1092 case BUILT_IN_ATOMIC_AND_FETCH_1:
1093 case BUILT_IN_ATOMIC_NAND_FETCH_1:
1094 case BUILT_IN_ATOMIC_XOR_FETCH_1:
1095 case BUILT_IN_ATOMIC_OR_FETCH_1:
1096 case BUILT_IN_ATOMIC_FETCH_ADD_1:
1097 case BUILT_IN_ATOMIC_FETCH_SUB_1:
1098 case BUILT_IN_ATOMIC_FETCH_AND_1:
1099 case BUILT_IN_ATOMIC_FETCH_NAND_1:
1100 case BUILT_IN_ATOMIC_FETCH_XOR_1:
1101 case BUILT_IN_ATOMIC_FETCH_OR_1:
1102 access_size = 1;
1103 goto do_atomic;
1104
1105 case BUILT_IN_ATOMIC_LOAD_2:
1106 is_store = false;
1107 /* FALLTHRU */
1108 case BUILT_IN_SYNC_FETCH_AND_ADD_2:
1109 case BUILT_IN_SYNC_FETCH_AND_SUB_2:
1110 case BUILT_IN_SYNC_FETCH_AND_OR_2:
1111 case BUILT_IN_SYNC_FETCH_AND_AND_2:
1112 case BUILT_IN_SYNC_FETCH_AND_XOR_2:
1113 case BUILT_IN_SYNC_FETCH_AND_NAND_2:
1114 case BUILT_IN_SYNC_ADD_AND_FETCH_2:
1115 case BUILT_IN_SYNC_SUB_AND_FETCH_2:
1116 case BUILT_IN_SYNC_OR_AND_FETCH_2:
1117 case BUILT_IN_SYNC_AND_AND_FETCH_2:
1118 case BUILT_IN_SYNC_XOR_AND_FETCH_2:
1119 case BUILT_IN_SYNC_NAND_AND_FETCH_2:
1120 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_2:
1121 case BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_2:
1122 case BUILT_IN_SYNC_LOCK_TEST_AND_SET_2:
1123 case BUILT_IN_SYNC_LOCK_RELEASE_2:
1124 case BUILT_IN_ATOMIC_EXCHANGE_2:
1125 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_2:
1126 case BUILT_IN_ATOMIC_STORE_2:
1127 case BUILT_IN_ATOMIC_ADD_FETCH_2:
1128 case BUILT_IN_ATOMIC_SUB_FETCH_2:
1129 case BUILT_IN_ATOMIC_AND_FETCH_2:
1130 case BUILT_IN_ATOMIC_NAND_FETCH_2:
1131 case BUILT_IN_ATOMIC_XOR_FETCH_2:
1132 case BUILT_IN_ATOMIC_OR_FETCH_2:
1133 case BUILT_IN_ATOMIC_FETCH_ADD_2:
1134 case BUILT_IN_ATOMIC_FETCH_SUB_2:
1135 case BUILT_IN_ATOMIC_FETCH_AND_2:
1136 case BUILT_IN_ATOMIC_FETCH_NAND_2:
1137 case BUILT_IN_ATOMIC_FETCH_XOR_2:
1138 case BUILT_IN_ATOMIC_FETCH_OR_2:
1139 access_size = 2;
1140 goto do_atomic;
1141
1142 case BUILT_IN_ATOMIC_LOAD_4:
1143 is_store = false;
1144 /* FALLTHRU */
1145 case BUILT_IN_SYNC_FETCH_AND_ADD_4:
1146 case BUILT_IN_SYNC_FETCH_AND_SUB_4:
1147 case BUILT_IN_SYNC_FETCH_AND_OR_4:
1148 case BUILT_IN_SYNC_FETCH_AND_AND_4:
1149 case BUILT_IN_SYNC_FETCH_AND_XOR_4:
1150 case BUILT_IN_SYNC_FETCH_AND_NAND_4:
1151 case BUILT_IN_SYNC_ADD_AND_FETCH_4:
1152 case BUILT_IN_SYNC_SUB_AND_FETCH_4:
1153 case BUILT_IN_SYNC_OR_AND_FETCH_4:
1154 case BUILT_IN_SYNC_AND_AND_FETCH_4:
1155 case BUILT_IN_SYNC_XOR_AND_FETCH_4:
1156 case BUILT_IN_SYNC_NAND_AND_FETCH_4:
1157 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_4:
1158 case BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_4:
1159 case BUILT_IN_SYNC_LOCK_TEST_AND_SET_4:
1160 case BUILT_IN_SYNC_LOCK_RELEASE_4:
1161 case BUILT_IN_ATOMIC_EXCHANGE_4:
1162 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_4:
1163 case BUILT_IN_ATOMIC_STORE_4:
1164 case BUILT_IN_ATOMIC_ADD_FETCH_4:
1165 case BUILT_IN_ATOMIC_SUB_FETCH_4:
1166 case BUILT_IN_ATOMIC_AND_FETCH_4:
1167 case BUILT_IN_ATOMIC_NAND_FETCH_4:
1168 case BUILT_IN_ATOMIC_XOR_FETCH_4:
1169 case BUILT_IN_ATOMIC_OR_FETCH_4:
1170 case BUILT_IN_ATOMIC_FETCH_ADD_4:
1171 case BUILT_IN_ATOMIC_FETCH_SUB_4:
1172 case BUILT_IN_ATOMIC_FETCH_AND_4:
1173 case BUILT_IN_ATOMIC_FETCH_NAND_4:
1174 case BUILT_IN_ATOMIC_FETCH_XOR_4:
1175 case BUILT_IN_ATOMIC_FETCH_OR_4:
1176 access_size = 4;
1177 goto do_atomic;
1178
1179 case BUILT_IN_ATOMIC_LOAD_8:
1180 is_store = false;
1181 /* FALLTHRU */
1182 case BUILT_IN_SYNC_FETCH_AND_ADD_8:
1183 case BUILT_IN_SYNC_FETCH_AND_SUB_8:
1184 case BUILT_IN_SYNC_FETCH_AND_OR_8:
1185 case BUILT_IN_SYNC_FETCH_AND_AND_8:
1186 case BUILT_IN_SYNC_FETCH_AND_XOR_8:
1187 case BUILT_IN_SYNC_FETCH_AND_NAND_8:
1188 case BUILT_IN_SYNC_ADD_AND_FETCH_8:
1189 case BUILT_IN_SYNC_SUB_AND_FETCH_8:
1190 case BUILT_IN_SYNC_OR_AND_FETCH_8:
1191 case BUILT_IN_SYNC_AND_AND_FETCH_8:
1192 case BUILT_IN_SYNC_XOR_AND_FETCH_8:
1193 case BUILT_IN_SYNC_NAND_AND_FETCH_8:
1194 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_8:
1195 case BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_8:
1196 case BUILT_IN_SYNC_LOCK_TEST_AND_SET_8:
1197 case BUILT_IN_SYNC_LOCK_RELEASE_8:
1198 case BUILT_IN_ATOMIC_EXCHANGE_8:
1199 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_8:
1200 case BUILT_IN_ATOMIC_STORE_8:
1201 case BUILT_IN_ATOMIC_ADD_FETCH_8:
1202 case BUILT_IN_ATOMIC_SUB_FETCH_8:
1203 case BUILT_IN_ATOMIC_AND_FETCH_8:
1204 case BUILT_IN_ATOMIC_NAND_FETCH_8:
1205 case BUILT_IN_ATOMIC_XOR_FETCH_8:
1206 case BUILT_IN_ATOMIC_OR_FETCH_8:
1207 case BUILT_IN_ATOMIC_FETCH_ADD_8:
1208 case BUILT_IN_ATOMIC_FETCH_SUB_8:
1209 case BUILT_IN_ATOMIC_FETCH_AND_8:
1210 case BUILT_IN_ATOMIC_FETCH_NAND_8:
1211 case BUILT_IN_ATOMIC_FETCH_XOR_8:
1212 case BUILT_IN_ATOMIC_FETCH_OR_8:
1213 access_size = 8;
1214 goto do_atomic;
1215
1216 case BUILT_IN_ATOMIC_LOAD_16:
1217 is_store = false;
1218 /* FALLTHRU */
1219 case BUILT_IN_SYNC_FETCH_AND_ADD_16:
1220 case BUILT_IN_SYNC_FETCH_AND_SUB_16:
1221 case BUILT_IN_SYNC_FETCH_AND_OR_16:
1222 case BUILT_IN_SYNC_FETCH_AND_AND_16:
1223 case BUILT_IN_SYNC_FETCH_AND_XOR_16:
1224 case BUILT_IN_SYNC_FETCH_AND_NAND_16:
1225 case BUILT_IN_SYNC_ADD_AND_FETCH_16:
1226 case BUILT_IN_SYNC_SUB_AND_FETCH_16:
1227 case BUILT_IN_SYNC_OR_AND_FETCH_16:
1228 case BUILT_IN_SYNC_AND_AND_FETCH_16:
1229 case BUILT_IN_SYNC_XOR_AND_FETCH_16:
1230 case BUILT_IN_SYNC_NAND_AND_FETCH_16:
1231 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_16:
1232 case BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_16:
1233 case BUILT_IN_SYNC_LOCK_TEST_AND_SET_16:
1234 case BUILT_IN_SYNC_LOCK_RELEASE_16:
1235 case BUILT_IN_ATOMIC_EXCHANGE_16:
1236 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_16:
1237 case BUILT_IN_ATOMIC_STORE_16:
1238 case BUILT_IN_ATOMIC_ADD_FETCH_16:
1239 case BUILT_IN_ATOMIC_SUB_FETCH_16:
1240 case BUILT_IN_ATOMIC_AND_FETCH_16:
1241 case BUILT_IN_ATOMIC_NAND_FETCH_16:
1242 case BUILT_IN_ATOMIC_XOR_FETCH_16:
1243 case BUILT_IN_ATOMIC_OR_FETCH_16:
1244 case BUILT_IN_ATOMIC_FETCH_ADD_16:
1245 case BUILT_IN_ATOMIC_FETCH_SUB_16:
1246 case BUILT_IN_ATOMIC_FETCH_AND_16:
1247 case BUILT_IN_ATOMIC_FETCH_NAND_16:
1248 case BUILT_IN_ATOMIC_FETCH_XOR_16:
1249 case BUILT_IN_ATOMIC_FETCH_OR_16:
1250 access_size = 16;
1251 /* FALLTHRU */
1252 do_atomic:
1253 {
1254 dest = gimple_call_arg (gs: call, index: 0);
1255 /* DEST represents the address of a memory location.
1256 instrument_derefs wants the memory location, so lets
1257 dereference the address DEST before handing it to
1258 instrument_derefs. */
1259 tree type = build_nonstandard_integer_type (access_size
1260 * BITS_PER_UNIT, 1);
1261 dest = build2 (MEM_REF, type, dest,
1262 build_int_cst (build_pointer_type (char_type_node), 0));
1263 break;
1264 }
1265
1266 default:
1267 /* The other builtins memory access are not instrumented in this
1268 function because they either don't have any length parameter,
1269 or their length parameter is just a limit. */
1270 break;
1271 }
1272
1273 if (len != NULL_TREE)
1274 {
1275 if (source0 != NULL_TREE)
1276 {
1277 src0->start = source0;
1278 src0->access_size = access_size;
1279 *src0_len = len;
1280 *src0_is_store = false;
1281 }
1282
1283 if (source1 != NULL_TREE)
1284 {
1285 src1->start = source1;
1286 src1->access_size = access_size;
1287 *src1_len = len;
1288 *src1_is_store = false;
1289 }
1290
1291 if (dest != NULL_TREE)
1292 {
1293 dst->start = dest;
1294 dst->access_size = access_size;
1295 *dst_len = len;
1296 *dst_is_store = true;
1297 }
1298
1299 got_reference_p = true;
1300 }
1301 else if (dest)
1302 {
1303 dst->start = dest;
1304 dst->access_size = access_size;
1305 *dst_len = NULL_TREE;
1306 *dst_is_store = is_store;
1307 *dest_is_deref = true;
1308 got_reference_p = true;
1309 }
1310
1311 return got_reference_p;
1312}
1313
1314/* Return true iff a given gimple statement has been instrumented.
1315 Note that the statement is "defined" by the memory references it
1316 contains. */
1317
1318static bool
1319has_stmt_been_instrumented_p (gimple *stmt)
1320{
1321 if (gimple_assign_single_p (gs: stmt))
1322 {
1323 bool r_is_store;
1324 asan_mem_ref r;
1325 asan_mem_ref_init (ref: &r, NULL, access_size: 1);
1326
1327 if (get_mem_ref_of_assignment (assignment: as_a <gassign *> (p: stmt), ref: &r,
1328 ref_is_store: &r_is_store))
1329 {
1330 if (!has_mem_ref_been_instrumented (ref: &r))
1331 return false;
1332 if (r_is_store && gimple_assign_load_p (stmt))
1333 {
1334 asan_mem_ref src;
1335 asan_mem_ref_init (ref: &src, NULL, access_size: 1);
1336 src.start = gimple_assign_rhs1 (gs: stmt);
1337 src.access_size = int_size_in_bytes (TREE_TYPE (src.start));
1338 if (!has_mem_ref_been_instrumented (ref: &src))
1339 return false;
1340 }
1341 return true;
1342 }
1343 }
1344 else if (gimple_call_builtin_p (stmt, BUILT_IN_NORMAL))
1345 {
1346 asan_mem_ref src0, src1, dest;
1347 asan_mem_ref_init (ref: &src0, NULL, access_size: 1);
1348 asan_mem_ref_init (ref: &src1, NULL, access_size: 1);
1349 asan_mem_ref_init (ref: &dest, NULL, access_size: 1);
1350
1351 tree src0_len = NULL_TREE, src1_len = NULL_TREE, dest_len = NULL_TREE;
1352 bool src0_is_store = false, src1_is_store = false,
1353 dest_is_store = false, dest_is_deref = false, intercepted_p = true;
1354 if (get_mem_refs_of_builtin_call (call: as_a <gcall *> (p: stmt),
1355 src0: &src0, src0_len: &src0_len, src0_is_store: &src0_is_store,
1356 src1: &src1, src1_len: &src1_len, src1_is_store: &src1_is_store,
1357 dst: &dest, dst_len: &dest_len, dst_is_store: &dest_is_store,
1358 dest_is_deref: &dest_is_deref, intercepted_p: &intercepted_p))
1359 {
1360 if (src0.start != NULL_TREE
1361 && !has_mem_ref_been_instrumented (ref: &src0, len: src0_len))
1362 return false;
1363
1364 if (src1.start != NULL_TREE
1365 && !has_mem_ref_been_instrumented (ref: &src1, len: src1_len))
1366 return false;
1367
1368 if (dest.start != NULL_TREE
1369 && !has_mem_ref_been_instrumented (ref: &dest, len: dest_len))
1370 return false;
1371
1372 return true;
1373 }
1374 }
1375 else if (is_gimple_call (gs: stmt)
1376 && gimple_store_p (gs: stmt)
1377 && (gimple_call_builtin_p (stmt)
1378 || gimple_call_internal_p (gs: stmt)
1379 || !aggregate_value_p (TREE_TYPE (gimple_call_lhs (stmt)),
1380 gimple_call_fntype (gs: stmt))))
1381 {
1382 asan_mem_ref r;
1383 asan_mem_ref_init (ref: &r, NULL, access_size: 1);
1384
1385 r.start = gimple_call_lhs (gs: stmt);
1386 r.access_size = int_size_in_bytes (TREE_TYPE (r.start));
1387 return has_mem_ref_been_instrumented (ref: &r);
1388 }
1389
1390 return false;
1391}
1392
1393/* Insert a memory reference into the hash table. */
1394
1395static void
1396update_mem_ref_hash_table (tree ref, HOST_WIDE_INT access_size)
1397{
1398 hash_table<asan_mem_ref_hasher> *ht = get_mem_ref_hash_table ();
1399
1400 asan_mem_ref r;
1401 asan_mem_ref_init (ref: &r, start: ref, access_size);
1402
1403 asan_mem_ref **slot = ht->find_slot (value: &r, insert: INSERT);
1404 if (*slot == NULL || (*slot)->access_size < access_size)
1405 *slot = asan_mem_ref_new (start: ref, access_size);
1406}
1407
1408/* Initialize shadow_ptr_types array. */
1409
1410static void
1411asan_init_shadow_ptr_types (void)
1412{
1413 asan_shadow_set = new_alias_set ();
1414 tree types[3] = { signed_char_type_node, short_integer_type_node,
1415 integer_type_node };
1416
1417 for (unsigned i = 0; i < 3; i++)
1418 {
1419 shadow_ptr_types[i] = build_distinct_type_copy (types[i]);
1420 TYPE_ALIAS_SET (shadow_ptr_types[i]) = asan_shadow_set;
1421 shadow_ptr_types[i] = build_pointer_type (shadow_ptr_types[i]);
1422 }
1423
1424 initialize_sanitizer_builtins ();
1425}
1426
1427/* Create ADDR_EXPR of STRING_CST with the PP pretty printer text. */
1428
1429static tree
1430asan_pp_string (pretty_printer *pp)
1431{
1432 const char *buf = pp_formatted_text (pp);
1433 size_t len = strlen (s: buf);
1434 tree ret = build_string (len + 1, buf);
1435 TREE_TYPE (ret)
1436 = build_array_type (TREE_TYPE (shadow_ptr_types[0]),
1437 build_index_type (size_int (len)));
1438 TREE_READONLY (ret) = 1;
1439 TREE_STATIC (ret) = 1;
1440 return build1 (ADDR_EXPR, shadow_ptr_types[0], ret);
1441}
1442
1443/* Clear shadow memory at SHADOW_MEM, LEN bytes. Can't call a library call here
1444 though. */
1445
1446static void
1447asan_clear_shadow (rtx shadow_mem, HOST_WIDE_INT len)
1448{
1449 rtx_insn *insn, *insns, *jump;
1450 rtx_code_label *top_label;
1451 rtx end, addr, tmp;
1452
1453 gcc_assert ((len & 3) == 0);
1454 start_sequence ();
1455 clear_storage (shadow_mem, GEN_INT (len), BLOCK_OP_NORMAL);
1456 insns = get_insns ();
1457 end_sequence ();
1458 for (insn = insns; insn; insn = NEXT_INSN (insn))
1459 if (CALL_P (insn))
1460 break;
1461 if (insn == NULL_RTX)
1462 {
1463 emit_insn (insns);
1464 return;
1465 }
1466
1467 top_label = gen_label_rtx ();
1468 addr = copy_to_mode_reg (Pmode, XEXP (shadow_mem, 0));
1469 shadow_mem = adjust_automodify_address (shadow_mem, SImode, addr, 0);
1470 end = force_reg (Pmode, plus_constant (Pmode, addr, len));
1471 emit_label (top_label);
1472
1473 emit_move_insn (shadow_mem, const0_rtx);
1474 tmp = expand_simple_binop (Pmode, PLUS, addr, gen_int_mode (4, Pmode), addr,
1475 true, OPTAB_LIB_WIDEN);
1476 if (tmp != addr)
1477 emit_move_insn (addr, tmp);
1478 emit_cmp_and_jump_insns (addr, end, LT, NULL_RTX, Pmode, true, top_label);
1479 jump = get_last_insn ();
1480 gcc_assert (JUMP_P (jump));
1481 add_reg_br_prob_note (jump,
1482 profile_probability::guessed_always ()
1483 .apply_scale (num: 80, den: 100));
1484}
1485
1486void
1487asan_function_start (void)
1488{
1489 ASM_OUTPUT_DEBUG_LABEL (asm_out_file, "LASANPC", current_function_funcdef_no);
1490}
1491
1492/* Return number of shadow bytes that are occupied by a local variable
1493 of SIZE bytes. */
1494
1495static unsigned HOST_WIDE_INT
1496shadow_mem_size (unsigned HOST_WIDE_INT size)
1497{
1498 /* It must be possible to align stack variables to granularity
1499 of shadow memory. */
1500 gcc_assert (BITS_PER_UNIT
1501 * ASAN_SHADOW_GRANULARITY <= MAX_SUPPORTED_STACK_ALIGNMENT);
1502
1503 return ROUND_UP (size, ASAN_SHADOW_GRANULARITY) / ASAN_SHADOW_GRANULARITY;
1504}
1505
1506/* Always emit 4 bytes at a time. */
1507#define RZ_BUFFER_SIZE 4
1508
1509/* ASAN redzone buffer container that handles emission of shadow bytes. */
1510class asan_redzone_buffer
1511{
1512public:
1513 /* Constructor. */
1514 asan_redzone_buffer (rtx shadow_mem, HOST_WIDE_INT prev_offset):
1515 m_shadow_mem (shadow_mem), m_prev_offset (prev_offset),
1516 m_original_offset (prev_offset), m_shadow_bytes (RZ_BUFFER_SIZE)
1517 {}
1518
1519 /* Emit VALUE shadow byte at a given OFFSET. */
1520 void emit_redzone_byte (HOST_WIDE_INT offset, unsigned char value);
1521
1522 /* Emit RTX emission of the content of the buffer. */
1523 void flush_redzone_payload (void);
1524
1525private:
1526 /* Flush if the content of the buffer is full
1527 (equal to RZ_BUFFER_SIZE). */
1528 void flush_if_full (void);
1529
1530 /* Memory where we last emitted a redzone payload. */
1531 rtx m_shadow_mem;
1532
1533 /* Relative offset where we last emitted a redzone payload. */
1534 HOST_WIDE_INT m_prev_offset;
1535
1536 /* Relative original offset. Used for checking only. */
1537 HOST_WIDE_INT m_original_offset;
1538
1539public:
1540 /* Buffer with redzone payload. */
1541 auto_vec<unsigned char> m_shadow_bytes;
1542};
1543
1544/* Emit VALUE shadow byte at a given OFFSET. */
1545
1546void
1547asan_redzone_buffer::emit_redzone_byte (HOST_WIDE_INT offset,
1548 unsigned char value)
1549{
1550 gcc_assert ((offset & (ASAN_SHADOW_GRANULARITY - 1)) == 0);
1551 gcc_assert (offset >= m_prev_offset);
1552
1553 HOST_WIDE_INT off
1554 = m_prev_offset + ASAN_SHADOW_GRANULARITY * m_shadow_bytes.length ();
1555 if (off == offset)
1556 /* Consecutive shadow memory byte. */;
1557 else if (offset < m_prev_offset + (HOST_WIDE_INT) (ASAN_SHADOW_GRANULARITY
1558 * RZ_BUFFER_SIZE)
1559 && !m_shadow_bytes.is_empty ())
1560 {
1561 /* Shadow memory byte with a small gap. */
1562 for (; off < offset; off += ASAN_SHADOW_GRANULARITY)
1563 m_shadow_bytes.safe_push (obj: 0);
1564 }
1565 else
1566 {
1567 if (!m_shadow_bytes.is_empty ())
1568 flush_redzone_payload ();
1569
1570 /* Maybe start earlier in order to use aligned store. */
1571 HOST_WIDE_INT align = (offset - m_prev_offset) % ASAN_RED_ZONE_SIZE;
1572 if (align)
1573 {
1574 offset -= align;
1575 for (unsigned i = 0; i < align / BITS_PER_UNIT; i++)
1576 m_shadow_bytes.safe_push (obj: 0);
1577 }
1578
1579 /* Adjust m_prev_offset and m_shadow_mem. */
1580 HOST_WIDE_INT diff = offset - m_prev_offset;
1581 m_shadow_mem = adjust_address (m_shadow_mem, VOIDmode,
1582 diff >> ASAN_SHADOW_SHIFT);
1583 m_prev_offset = offset;
1584 }
1585 m_shadow_bytes.safe_push (obj: value);
1586 flush_if_full ();
1587}
1588
1589/* Emit RTX emission of the content of the buffer. */
1590
1591void
1592asan_redzone_buffer::flush_redzone_payload (void)
1593{
1594 gcc_assert (WORDS_BIG_ENDIAN == BYTES_BIG_ENDIAN);
1595
1596 if (m_shadow_bytes.is_empty ())
1597 return;
1598
1599 /* Be sure we always emit to an aligned address. */
1600 gcc_assert (((m_prev_offset - m_original_offset)
1601 & (ASAN_RED_ZONE_SIZE - 1)) == 0);
1602
1603 /* Fill it to RZ_BUFFER_SIZE bytes with zeros if needed. */
1604 unsigned l = m_shadow_bytes.length ();
1605 for (unsigned i = 0; i <= RZ_BUFFER_SIZE - l; i++)
1606 m_shadow_bytes.safe_push (obj: 0);
1607
1608 if (dump_file && (dump_flags & TDF_DETAILS))
1609 fprintf (stream: dump_file,
1610 format: "Flushing rzbuffer at offset %" PRId64 " with: ", m_prev_offset);
1611
1612 unsigned HOST_WIDE_INT val = 0;
1613 for (unsigned i = 0; i < RZ_BUFFER_SIZE; i++)
1614 {
1615 unsigned char v
1616 = m_shadow_bytes[BYTES_BIG_ENDIAN ? RZ_BUFFER_SIZE - i - 1 : i];
1617 val |= (unsigned HOST_WIDE_INT)v << (BITS_PER_UNIT * i);
1618 if (dump_file && (dump_flags & TDF_DETAILS))
1619 fprintf (stream: dump_file, format: "%02x ", v);
1620 }
1621
1622 if (dump_file && (dump_flags & TDF_DETAILS))
1623 fprintf (stream: dump_file, format: "\n");
1624
1625 rtx c = gen_int_mode (val, SImode);
1626 m_shadow_mem = adjust_address (m_shadow_mem, SImode, 0);
1627 emit_move_insn (m_shadow_mem, c);
1628 m_shadow_bytes.truncate (size: 0);
1629}
1630
1631/* Flush if the content of the buffer is full
1632 (equal to RZ_BUFFER_SIZE). */
1633
1634void
1635asan_redzone_buffer::flush_if_full (void)
1636{
1637 if (m_shadow_bytes.length () == RZ_BUFFER_SIZE)
1638 flush_redzone_payload ();
1639}
1640
1641
1642/* HWAddressSanitizer (hwasan) is a probabilistic method for detecting
1643 out-of-bounds and use-after-free bugs.
1644 Read more:
1645 http://code.google.com/p/address-sanitizer/
1646
1647 Similar to AddressSanitizer (asan) it consists of two parts: the
1648 instrumentation module in this file, and a run-time library.
1649
1650 The instrumentation module adds a run-time check before every memory insn in
1651 the same manner as asan (see the block comment for AddressSanitizer above).
1652 Currently, hwasan only adds out-of-line instrumentation, where each check is
1653 implemented as a function call to the run-time library. Hence a check for a
1654 load of N bytes from address X would be implemented with a function call to
1655 __hwasan_loadN(X), and checking a store of N bytes from address X would be
1656 implemented with a function call to __hwasan_storeN(X).
1657
1658 The main difference between hwasan and asan is in the information stored to
1659 help this checking. Both sanitizers use a shadow memory area which stores
1660 data recording the state of main memory at a corresponding address.
1661
1662 For hwasan, each 16 byte granule in main memory has a corresponding 1 byte
1663 in shadow memory. This shadow address can be calculated with equation:
1664 (addr >> log_2(HWASAN_TAG_GRANULE_SIZE))
1665 + __hwasan_shadow_memory_dynamic_address;
1666 The conversion between real and shadow memory for asan is given in the block
1667 comment at the top of this file.
1668 The description of how this shadow memory is laid out for asan is in the
1669 block comment at the top of this file, here we describe how this shadow
1670 memory is used for hwasan.
1671
1672 For hwasan, each variable is assigned a byte-sized 'tag'. The extent of
1673 the shadow memory for that variable is filled with the assigned tag, and
1674 every pointer referencing that variable has its top byte set to the same
1675 tag. The run-time library redefines malloc so that every allocation returns
1676 a tagged pointer and tags the corresponding shadow memory with the same tag.
1677
1678 On each pointer dereference the tag found in the pointer is compared to the
1679 tag found in the shadow memory corresponding to the accessed memory address.
1680 If these tags are found to differ then this memory access is judged to be
1681 invalid and a report is generated.
1682
1683 This method of bug detection is not perfect -- it can not catch every bad
1684 access -- but catches them probabilistically instead. There is always the
1685 possibility that an invalid memory access will happen to access memory
1686 tagged with the same tag as the pointer that this access used.
1687 The chances of this are approx. 0.4% for any two uncorrelated objects.
1688
1689 Random tag generation can mitigate this problem by decreasing the
1690 probability that an invalid access will be missed in the same manner over
1691 multiple runs. i.e. if two objects are tagged the same in one run of the
1692 binary they are unlikely to be tagged the same in the next run.
1693 Both heap and stack allocated objects have random tags by default.
1694
1695 [16 byte granule implications]
1696 Since the shadow memory only has a resolution on real memory of 16 bytes,
1697 invalid accesses that are within the same 16 byte granule as a valid
1698 address will not be caught.
1699
1700 There is a "short-granule" feature in the runtime library which does catch
1701 such accesses, but this feature is not implemented for stack objects (since
1702 stack objects are allocated and tagged by compiler instrumentation, and
1703 this feature has not yet been implemented in GCC instrumentation).
1704
1705 Another outcome of this 16 byte resolution is that each tagged object must
1706 be 16 byte aligned. If two objects were to share any 16 byte granule in
1707 memory, then they both would have to be given the same tag, and invalid
1708 accesses to one using a pointer to the other would be undetectable.
1709
1710 [Compiler instrumentation]
1711 Compiler instrumentation ensures that two adjacent buffers on the stack are
1712 given different tags, this means an access to one buffer using a pointer
1713 generated from the other (e.g. through buffer overrun) will have mismatched
1714 tags and be caught by hwasan.
1715
1716 We don't randomly tag every object on the stack, since that would require
1717 keeping many registers to record each tag. Instead we randomly generate a
1718 tag for each function frame, and each new stack object uses a tag offset
1719 from that frame tag.
1720 i.e. each object is tagged as RFT + offset, where RFT is the "random frame
1721 tag" generated for this frame.
1722 This means that randomisation does not peturb the difference between tags
1723 on tagged stack objects within a frame, but this is mitigated by the fact
1724 that objects with the same tag within a frame are very far apart
1725 (approx. 2^HWASAN_TAG_SIZE objects apart).
1726
1727 As a demonstration, using the same example program as in the asan block
1728 comment above:
1729
1730 int
1731 foo ()
1732 {
1733 char a[24] = {0};
1734 int b[2] = {0};
1735
1736 a[5] = 1;
1737 b[1] = 2;
1738
1739 return a[5] + b[1];
1740 }
1741
1742 On AArch64 the stack will be ordered as follows for the above function:
1743
1744 Slot 1/ [24 bytes for variable 'a']
1745 Slot 2/ [8 bytes padding for alignment]
1746 Slot 3/ [8 bytes for variable 'b']
1747 Slot 4/ [8 bytes padding for alignment]
1748
1749 (The padding is there to ensure 16 byte alignment as described in the 16
1750 byte granule implications).
1751
1752 While the shadow memory will be ordered as follows:
1753
1754 - 2 bytes (representing 32 bytes in real memory) tagged with RFT + 1.
1755 - 1 byte (representing 16 bytes in real memory) tagged with RFT + 2.
1756
1757 And any pointer to "a" will have the tag RFT + 1, and any pointer to "b"
1758 will have the tag RFT + 2.
1759
1760 [Top Byte Ignore requirements]
1761 Hwasan requires the ability to store an 8 bit tag in every pointer. There
1762 is no instrumentation done to remove this tag from pointers before
1763 dereferencing, which means the hardware must ignore this tag during memory
1764 accesses.
1765
1766 Architectures where this feature is available should indicate this using
1767 the TARGET_MEMTAG_CAN_TAG_ADDRESSES hook.
1768
1769 [Stack requires cleanup on unwinding]
1770 During normal operation of a hwasan sanitized program more space in the
1771 shadow memory becomes tagged as the stack grows. As the stack shrinks this
1772 shadow memory space must become untagged. If it is not untagged then when
1773 the stack grows again (during other function calls later on in the program)
1774 objects on the stack that are usually not tagged (e.g. parameters passed on
1775 the stack) can be placed in memory whose shadow space is tagged with
1776 something else, and accesses can cause false positive reports.
1777
1778 Hence we place untagging code on every epilogue of functions which tag some
1779 stack objects.
1780
1781 Moreover, the run-time library intercepts longjmp & setjmp to untag when
1782 the stack is unwound this way.
1783
1784 C++ exceptions are not yet handled, which means this sanitizer can not
1785 handle C++ code that throws exceptions -- it will give false positives
1786 after an exception has been thrown. The implementation that the hwasan
1787 library has for handling these relies on the frame pointer being after any
1788 local variables. This is not generally the case for GCC. */
1789
1790
1791/* Returns whether we are tagging pointers and checking those tags on memory
1792 access. */
1793bool
1794hwasan_sanitize_p ()
1795{
1796 return sanitize_flags_p (flag: SANITIZE_HWADDRESS);
1797}
1798
1799/* Are we tagging the stack? */
1800bool
1801hwasan_sanitize_stack_p ()
1802{
1803 return (hwasan_sanitize_p () && param_hwasan_instrument_stack);
1804}
1805
1806/* Are we tagging alloca objects? */
1807bool
1808hwasan_sanitize_allocas_p (void)
1809{
1810 return (hwasan_sanitize_stack_p () && param_hwasan_instrument_allocas);
1811}
1812
1813/* Should we instrument reads? */
1814bool
1815hwasan_instrument_reads (void)
1816{
1817 return (hwasan_sanitize_p () && param_hwasan_instrument_reads);
1818}
1819
1820/* Should we instrument writes? */
1821bool
1822hwasan_instrument_writes (void)
1823{
1824 return (hwasan_sanitize_p () && param_hwasan_instrument_writes);
1825}
1826
1827/* Should we instrument builtin calls? */
1828bool
1829hwasan_memintrin (void)
1830{
1831 return (hwasan_sanitize_p () && param_hwasan_instrument_mem_intrinsics);
1832}
1833
1834/* Insert code to protect stack vars. The prologue sequence should be emitted
1835 directly, epilogue sequence returned. BASE is the register holding the
1836 stack base, against which OFFSETS array offsets are relative to, OFFSETS
1837 array contains pairs of offsets in reverse order, always the end offset
1838 of some gap that needs protection followed by starting offset,
1839 and DECLS is an array of representative decls for each var partition.
1840 LENGTH is the length of the OFFSETS array, DECLS array is LENGTH / 2 - 1
1841 elements long (OFFSETS include gap before the first variable as well
1842 as gaps after each stack variable). PBASE is, if non-NULL, some pseudo
1843 register which stack vars DECL_RTLs are based on. Either BASE should be
1844 assigned to PBASE, when not doing use after return protection, or
1845 corresponding address based on __asan_stack_malloc* return value. */
1846
1847rtx_insn *
1848asan_emit_stack_protection (rtx base, rtx pbase, unsigned int alignb,
1849 HOST_WIDE_INT *offsets, tree *decls, int length)
1850{
1851 rtx shadow_base, shadow_mem, ret, mem, orig_base;
1852 rtx_code_label *lab;
1853 rtx_insn *insns;
1854 char buf[32];
1855 HOST_WIDE_INT base_offset = offsets[length - 1];
1856 HOST_WIDE_INT base_align_bias = 0, offset, prev_offset;
1857 HOST_WIDE_INT asan_frame_size = offsets[0] - base_offset;
1858 HOST_WIDE_INT last_offset, last_size, last_size_aligned;
1859 int l;
1860 unsigned char cur_shadow_byte = ASAN_STACK_MAGIC_LEFT;
1861 tree str_cst, decl, id;
1862 int use_after_return_class = -1;
1863
1864 /* Don't emit anything when doing error recovery, the assertions
1865 might fail e.g. if a function had a frame offset overflow. */
1866 if (seen_error ())
1867 return NULL;
1868
1869 if (shadow_ptr_types[0] == NULL_TREE)
1870 asan_init_shadow_ptr_types ();
1871
1872 expanded_location cfun_xloc
1873 = expand_location (DECL_SOURCE_LOCATION (current_function_decl));
1874
1875 /* First of all, prepare the description string. */
1876 pretty_printer asan_pp;
1877
1878 pp_decimal_int (&asan_pp, length / 2 - 1);
1879 pp_space (&asan_pp);
1880 for (l = length - 2; l; l -= 2)
1881 {
1882 tree decl = decls[l / 2 - 1];
1883 pp_wide_integer (pp: &asan_pp, i: offsets[l] - base_offset);
1884 pp_space (&asan_pp);
1885 pp_wide_integer (pp: &asan_pp, i: offsets[l - 1] - offsets[l]);
1886 pp_space (&asan_pp);
1887
1888 expanded_location xloc
1889 = expand_location (DECL_SOURCE_LOCATION (decl));
1890 char location[32];
1891
1892 if (xloc.file == cfun_xloc.file)
1893 sprintf (s: location, format: ":%d", xloc.line);
1894 else
1895 location[0] = '\0';
1896
1897 if (DECL_P (decl) && DECL_NAME (decl))
1898 {
1899 unsigned idlen
1900 = IDENTIFIER_LENGTH (DECL_NAME (decl)) + strlen (s: location);
1901 pp_decimal_int (&asan_pp, idlen);
1902 pp_space (&asan_pp);
1903 pp_tree_identifier (&asan_pp, DECL_NAME (decl));
1904 pp_string (&asan_pp, location);
1905 }
1906 else
1907 pp_string (&asan_pp, "9 <unknown>");
1908
1909 if (l > 2)
1910 pp_space (&asan_pp);
1911 }
1912 str_cst = asan_pp_string (pp: &asan_pp);
1913
1914 gcc_checking_assert (offsets[0] == (crtl->stack_protect_guard
1915 ? -ASAN_RED_ZONE_SIZE : 0));
1916 /* Emit the prologue sequence. */
1917 if (asan_frame_size > 32 && asan_frame_size <= 65536 && pbase
1918 && param_asan_use_after_return)
1919 {
1920 HOST_WIDE_INT adjusted_frame_size = asan_frame_size;
1921 /* The stack protector guard is allocated at the top of the frame
1922 and cfgexpand.cc then uses align_frame_offset (ASAN_RED_ZONE_SIZE);
1923 while in that case we can still use asan_frame_size, we need to take
1924 that into account when computing base_align_bias. */
1925 if (alignb > ASAN_RED_ZONE_SIZE && crtl->stack_protect_guard)
1926 adjusted_frame_size += ASAN_RED_ZONE_SIZE;
1927 use_after_return_class = floor_log2 (x: asan_frame_size - 1) - 5;
1928 /* __asan_stack_malloc_N guarantees alignment
1929 N < 6 ? (64 << N) : 4096 bytes. */
1930 if (alignb > (use_after_return_class < 6
1931 ? (64U << use_after_return_class) : 4096U))
1932 use_after_return_class = -1;
1933 else if (alignb > ASAN_RED_ZONE_SIZE
1934 && (adjusted_frame_size & (alignb - 1)))
1935 {
1936 base_align_bias
1937 = ((adjusted_frame_size + alignb - 1)
1938 & ~(alignb - HOST_WIDE_INT_1)) - adjusted_frame_size;
1939 use_after_return_class
1940 = floor_log2 (x: asan_frame_size + base_align_bias - 1) - 5;
1941 if (use_after_return_class > 10)
1942 {
1943 base_align_bias = 0;
1944 use_after_return_class = -1;
1945 }
1946 }
1947 }
1948
1949 /* Align base if target is STRICT_ALIGNMENT. */
1950 if (STRICT_ALIGNMENT)
1951 {
1952 const HOST_WIDE_INT align
1953 = (GET_MODE_ALIGNMENT (SImode) / BITS_PER_UNIT) << ASAN_SHADOW_SHIFT;
1954 base = expand_binop (Pmode, and_optab, base, gen_int_mode (-align, Pmode),
1955 NULL_RTX, 1, OPTAB_DIRECT);
1956 }
1957
1958 if (use_after_return_class == -1 && pbase)
1959 emit_move_insn (pbase, base);
1960
1961 base = expand_binop (Pmode, add_optab, base,
1962 gen_int_mode (base_offset - base_align_bias, Pmode),
1963 NULL_RTX, 1, OPTAB_DIRECT);
1964 orig_base = NULL_RTX;
1965 if (use_after_return_class != -1)
1966 {
1967 if (asan_detect_stack_use_after_return == NULL_TREE)
1968 {
1969 id = get_identifier ("__asan_option_detect_stack_use_after_return");
1970 decl = build_decl (BUILTINS_LOCATION, VAR_DECL, id,
1971 integer_type_node);
1972 SET_DECL_ASSEMBLER_NAME (decl, id);
1973 TREE_ADDRESSABLE (decl) = 1;
1974 DECL_ARTIFICIAL (decl) = 1;
1975 DECL_IGNORED_P (decl) = 1;
1976 DECL_EXTERNAL (decl) = 1;
1977 TREE_STATIC (decl) = 1;
1978 TREE_PUBLIC (decl) = 1;
1979 TREE_USED (decl) = 1;
1980 asan_detect_stack_use_after_return = decl;
1981 }
1982 orig_base = gen_reg_rtx (Pmode);
1983 emit_move_insn (orig_base, base);
1984 ret = expand_normal (exp: asan_detect_stack_use_after_return);
1985 lab = gen_label_rtx ();
1986 emit_cmp_and_jump_insns (ret, const0_rtx, EQ, NULL_RTX,
1987 VOIDmode, 0, lab,
1988 prob: profile_probability::very_likely ());
1989 snprintf (s: buf, maxlen: sizeof buf, format: "__asan_stack_malloc_%d",
1990 use_after_return_class);
1991 ret = init_one_libfunc (buf);
1992 ret = emit_library_call_value (fun: ret, NULL_RTX, fn_type: LCT_NORMAL, outmode: ptr_mode,
1993 GEN_INT (asan_frame_size
1994 + base_align_bias),
1995 TYPE_MODE (pointer_sized_int_node));
1996 /* __asan_stack_malloc_[n] returns a pointer to fake stack if succeeded
1997 and NULL otherwise. Check RET value is NULL here and jump over the
1998 BASE reassignment in this case. Otherwise, reassign BASE to RET. */
1999 emit_cmp_and_jump_insns (ret, const0_rtx, EQ, NULL_RTX,
2000 VOIDmode, 0, lab,
2001 prob: profile_probability:: very_unlikely ());
2002 ret = convert_memory_address (Pmode, ret);
2003 emit_move_insn (base, ret);
2004 emit_label (lab);
2005 emit_move_insn (pbase, expand_binop (Pmode, add_optab, base,
2006 gen_int_mode (base_align_bias
2007 - base_offset, Pmode),
2008 NULL_RTX, 1, OPTAB_DIRECT));
2009 }
2010 mem = gen_rtx_MEM (ptr_mode, base);
2011 mem = adjust_address (mem, VOIDmode, base_align_bias);
2012 emit_move_insn (mem, gen_int_mode (ASAN_STACK_FRAME_MAGIC, ptr_mode));
2013 mem = adjust_address (mem, VOIDmode, GET_MODE_SIZE (ptr_mode));
2014 emit_move_insn (mem, expand_normal (exp: str_cst));
2015 mem = adjust_address (mem, VOIDmode, GET_MODE_SIZE (ptr_mode));
2016 ASM_GENERATE_INTERNAL_LABEL (buf, "LASANPC", current_function_funcdef_no);
2017 id = get_identifier (buf);
2018 decl = build_decl (DECL_SOURCE_LOCATION (current_function_decl),
2019 VAR_DECL, id, char_type_node);
2020 SET_DECL_ASSEMBLER_NAME (decl, id);
2021 TREE_ADDRESSABLE (decl) = 1;
2022 TREE_READONLY (decl) = 1;
2023 DECL_ARTIFICIAL (decl) = 1;
2024 DECL_IGNORED_P (decl) = 1;
2025 TREE_STATIC (decl) = 1;
2026 TREE_PUBLIC (decl) = 0;
2027 TREE_USED (decl) = 1;
2028 DECL_INITIAL (decl) = decl;
2029 TREE_ASM_WRITTEN (decl) = 1;
2030 TREE_ASM_WRITTEN (id) = 1;
2031 DECL_ALIGN_RAW (decl) = DECL_ALIGN_RAW (current_function_decl);
2032 emit_move_insn (mem, expand_normal (build_fold_addr_expr (decl)));
2033 shadow_base = expand_binop (Pmode, lshr_optab, base,
2034 gen_int_shift_amount (Pmode, ASAN_SHADOW_SHIFT),
2035 NULL_RTX, 1, OPTAB_DIRECT);
2036 shadow_base
2037 = plus_constant (Pmode, shadow_base,
2038 asan_shadow_offset ()
2039 + (base_align_bias >> ASAN_SHADOW_SHIFT));
2040 gcc_assert (asan_shadow_set != -1
2041 && (ASAN_RED_ZONE_SIZE >> ASAN_SHADOW_SHIFT) == 4);
2042 shadow_mem = gen_rtx_MEM (SImode, shadow_base);
2043 set_mem_alias_set (shadow_mem, asan_shadow_set);
2044 if (STRICT_ALIGNMENT)
2045 set_mem_align (shadow_mem, (GET_MODE_ALIGNMENT (SImode)));
2046 prev_offset = base_offset;
2047
2048 asan_redzone_buffer rz_buffer (shadow_mem, prev_offset);
2049 for (l = length; l; l -= 2)
2050 {
2051 if (l == 2)
2052 cur_shadow_byte = ASAN_STACK_MAGIC_RIGHT;
2053 offset = offsets[l - 1];
2054
2055 bool extra_byte = (offset - base_offset) & (ASAN_SHADOW_GRANULARITY - 1);
2056 /* If a red-zone is not aligned to ASAN_SHADOW_GRANULARITY then
2057 the previous stack variable has size % ASAN_SHADOW_GRANULARITY != 0.
2058 In that case we have to emit one extra byte that will describe
2059 how many bytes (our of ASAN_SHADOW_GRANULARITY) can be accessed. */
2060 if (extra_byte)
2061 {
2062 HOST_WIDE_INT aoff
2063 = base_offset + ((offset - base_offset)
2064 & ~(ASAN_SHADOW_GRANULARITY - HOST_WIDE_INT_1));
2065 rz_buffer.emit_redzone_byte (offset: aoff, value: offset - aoff);
2066 offset = aoff + ASAN_SHADOW_GRANULARITY;
2067 }
2068
2069 /* Calculate size of red zone payload. */
2070 while (offset < offsets[l - 2])
2071 {
2072 rz_buffer.emit_redzone_byte (offset, value: cur_shadow_byte);
2073 offset += ASAN_SHADOW_GRANULARITY;
2074 }
2075
2076 cur_shadow_byte = ASAN_STACK_MAGIC_MIDDLE;
2077 }
2078
2079 /* As the automatic variables are aligned to
2080 ASAN_RED_ZONE_SIZE / ASAN_SHADOW_GRANULARITY, the buffer should be
2081 flushed here. */
2082 gcc_assert (rz_buffer.m_shadow_bytes.is_empty ());
2083
2084 do_pending_stack_adjust ();
2085
2086 /* Construct epilogue sequence. */
2087 start_sequence ();
2088
2089 lab = NULL;
2090 if (use_after_return_class != -1)
2091 {
2092 rtx_code_label *lab2 = gen_label_rtx ();
2093 char c = (char) ASAN_STACK_MAGIC_USE_AFTER_RET;
2094 emit_cmp_and_jump_insns (orig_base, base, EQ, NULL_RTX,
2095 VOIDmode, 0, lab2,
2096 prob: profile_probability::very_likely ());
2097 shadow_mem = gen_rtx_MEM (BLKmode, shadow_base);
2098 set_mem_alias_set (shadow_mem, asan_shadow_set);
2099 mem = gen_rtx_MEM (ptr_mode, base);
2100 mem = adjust_address (mem, VOIDmode, base_align_bias);
2101 emit_move_insn (mem, gen_int_mode (ASAN_STACK_RETIRED_MAGIC, ptr_mode));
2102 unsigned HOST_WIDE_INT sz = asan_frame_size >> ASAN_SHADOW_SHIFT;
2103 if (use_after_return_class < 5
2104 && can_store_by_pieces (sz, builtin_memset_read_str, &c,
2105 BITS_PER_UNIT, true))
2106 {
2107 /* Emit:
2108 memset(ShadowBase, kAsanStackAfterReturnMagic, ShadowSize);
2109 **SavedFlagPtr(FakeStack, class_id) = 0
2110 */
2111 store_by_pieces (shadow_mem, sz, builtin_memset_read_str, &c,
2112 BITS_PER_UNIT, true, RETURN_BEGIN);
2113
2114 unsigned HOST_WIDE_INT offset
2115 = (1 << (use_after_return_class + 6));
2116 offset -= GET_MODE_SIZE (mode: ptr_mode);
2117 mem = gen_rtx_MEM (ptr_mode, base);
2118 mem = adjust_address (mem, ptr_mode, offset);
2119 rtx addr = gen_reg_rtx (ptr_mode);
2120 emit_move_insn (addr, mem);
2121 addr = convert_memory_address (Pmode, addr);
2122 mem = gen_rtx_MEM (QImode, addr);
2123 emit_move_insn (mem, const0_rtx);
2124 }
2125 else if (use_after_return_class >= 5
2126 || !set_storage_via_setmem (shadow_mem,
2127 GEN_INT (sz),
2128 gen_int_mode (c, QImode),
2129 BITS_PER_UNIT, BITS_PER_UNIT,
2130 -1, sz, sz, sz))
2131 {
2132 snprintf (s: buf, maxlen: sizeof buf, format: "__asan_stack_free_%d",
2133 use_after_return_class);
2134 ret = init_one_libfunc (buf);
2135 rtx addr = convert_memory_address (ptr_mode, base);
2136 rtx orig_addr = convert_memory_address (ptr_mode, orig_base);
2137 emit_library_call (fun: ret, fn_type: LCT_NORMAL, outmode: ptr_mode, arg1: addr, arg1_mode: ptr_mode,
2138 GEN_INT (asan_frame_size + base_align_bias),
2139 TYPE_MODE (pointer_sized_int_node),
2140 arg3: orig_addr, arg3_mode: ptr_mode);
2141 }
2142 lab = gen_label_rtx ();
2143 emit_jump (lab);
2144 emit_label (lab2);
2145 }
2146
2147 shadow_mem = gen_rtx_MEM (BLKmode, shadow_base);
2148 set_mem_alias_set (shadow_mem, asan_shadow_set);
2149
2150 if (STRICT_ALIGNMENT)
2151 set_mem_align (shadow_mem, (GET_MODE_ALIGNMENT (SImode)));
2152
2153 prev_offset = base_offset;
2154 last_offset = base_offset;
2155 last_size = 0;
2156 last_size_aligned = 0;
2157 for (l = length; l; l -= 2)
2158 {
2159 offset = base_offset + ((offsets[l - 1] - base_offset)
2160 & ~(ASAN_RED_ZONE_SIZE - HOST_WIDE_INT_1));
2161 if (last_offset + last_size_aligned < offset)
2162 {
2163 shadow_mem = adjust_address (shadow_mem, VOIDmode,
2164 (last_offset - prev_offset)
2165 >> ASAN_SHADOW_SHIFT);
2166 prev_offset = last_offset;
2167 asan_clear_shadow (shadow_mem, len: last_size_aligned >> ASAN_SHADOW_SHIFT);
2168 last_offset = offset;
2169 last_size = 0;
2170 }
2171 else
2172 last_size = offset - last_offset;
2173 last_size += base_offset + ((offsets[l - 2] - base_offset)
2174 & ~(ASAN_MIN_RED_ZONE_SIZE - HOST_WIDE_INT_1))
2175 - offset;
2176
2177 /* Unpoison shadow memory that corresponds to a variable that is
2178 is subject of use-after-return sanitization. */
2179 if (l > 2)
2180 {
2181 decl = decls[l / 2 - 2];
2182 if (asan_handled_variables != NULL
2183 && asan_handled_variables->contains (k: decl))
2184 {
2185 HOST_WIDE_INT size = offsets[l - 3] - offsets[l - 2];
2186 if (dump_file && (dump_flags & TDF_DETAILS))
2187 {
2188 const char *n = (DECL_NAME (decl)
2189 ? IDENTIFIER_POINTER (DECL_NAME (decl))
2190 : "<unknown>");
2191 fprintf (stream: dump_file, format: "Unpoisoning shadow stack for variable: "
2192 "%s (%" PRId64 " B)\n", n, size);
2193 }
2194
2195 last_size += size & ~(ASAN_MIN_RED_ZONE_SIZE - HOST_WIDE_INT_1);
2196 }
2197 }
2198 last_size_aligned
2199 = ((last_size + (ASAN_RED_ZONE_SIZE - HOST_WIDE_INT_1))
2200 & ~(ASAN_RED_ZONE_SIZE - HOST_WIDE_INT_1));
2201 }
2202 if (last_size_aligned)
2203 {
2204 shadow_mem = adjust_address (shadow_mem, VOIDmode,
2205 (last_offset - prev_offset)
2206 >> ASAN_SHADOW_SHIFT);
2207 asan_clear_shadow (shadow_mem, len: last_size_aligned >> ASAN_SHADOW_SHIFT);
2208 }
2209
2210 /* Clean-up set with instrumented stack variables. */
2211 delete asan_handled_variables;
2212 asan_handled_variables = NULL;
2213 delete asan_used_labels;
2214 asan_used_labels = NULL;
2215
2216 do_pending_stack_adjust ();
2217 if (lab)
2218 emit_label (lab);
2219
2220 insns = get_insns ();
2221 end_sequence ();
2222 return insns;
2223}
2224
2225/* Emit __asan_allocas_unpoison (top, bot) call. The BASE parameter corresponds
2226 to BOT argument, for TOP virtual_stack_dynamic_rtx is used. NEW_SEQUENCE
2227 indicates whether we're emitting new instructions sequence or not. */
2228
2229rtx_insn *
2230asan_emit_allocas_unpoison (rtx top, rtx bot, rtx_insn *before)
2231{
2232 if (before)
2233 push_to_sequence (before);
2234 else
2235 start_sequence ();
2236 rtx ret = init_one_libfunc ("__asan_allocas_unpoison");
2237 top = convert_memory_address (ptr_mode, top);
2238 bot = convert_memory_address (ptr_mode, bot);
2239 emit_library_call (fun: ret, fn_type: LCT_NORMAL, outmode: ptr_mode,
2240 arg1: top, arg1_mode: ptr_mode, arg2: bot, arg2_mode: ptr_mode);
2241
2242 do_pending_stack_adjust ();
2243 rtx_insn *insns = get_insns ();
2244 end_sequence ();
2245 return insns;
2246}
2247
2248/* Return true if DECL, a global var, might be overridden and needs
2249 therefore a local alias. */
2250
2251static bool
2252asan_needs_local_alias (tree decl)
2253{
2254 return DECL_WEAK (decl) || !targetm.binds_local_p (decl);
2255}
2256
2257/* Return true if DECL, a global var, is an artificial ODR indicator symbol
2258 therefore doesn't need protection. */
2259
2260static bool
2261is_odr_indicator (tree decl)
2262{
2263 return (DECL_ARTIFICIAL (decl)
2264 && lookup_attribute (attr_name: "asan odr indicator", DECL_ATTRIBUTES (decl)));
2265}
2266
2267/* Return true if DECL is a VAR_DECL that should be protected
2268 by Address Sanitizer, by appending a red zone with protected
2269 shadow memory after it and aligning it to at least
2270 ASAN_RED_ZONE_SIZE bytes. */
2271
2272bool
2273asan_protect_global (tree decl, bool ignore_decl_rtl_set_p)
2274{
2275 if (!param_asan_globals)
2276 return false;
2277
2278 rtx rtl, symbol;
2279
2280 if (TREE_CODE (decl) == STRING_CST)
2281 {
2282 /* Instrument all STRING_CSTs except those created
2283 by asan_pp_string here. */
2284 if (shadow_ptr_types[0] != NULL_TREE
2285 && TREE_CODE (TREE_TYPE (decl)) == ARRAY_TYPE
2286 && TREE_TYPE (TREE_TYPE (decl)) == TREE_TYPE (shadow_ptr_types[0]))
2287 return false;
2288 return true;
2289 }
2290 if (!VAR_P (decl)
2291 /* TLS vars aren't statically protectable. */
2292 || DECL_THREAD_LOCAL_P (decl)
2293 /* Externs will be protected elsewhere. */
2294 || DECL_EXTERNAL (decl)
2295 /* PR sanitizer/81697: For architectures that use section anchors first
2296 call to asan_protect_global may occur before DECL_RTL (decl) is set.
2297 We should ignore DECL_RTL_SET_P then, because otherwise the first call
2298 to asan_protect_global will return FALSE and the following calls on the
2299 same decl after setting DECL_RTL (decl) will return TRUE and we'll end
2300 up with inconsistency at runtime. */
2301 || (!DECL_RTL_SET_P (decl) && !ignore_decl_rtl_set_p)
2302 /* Comdat vars pose an ABI problem, we can't know if
2303 the var that is selected by the linker will have
2304 padding or not. */
2305 || DECL_ONE_ONLY (decl)
2306 /* Similarly for common vars. People can use -fno-common.
2307 Note: Linux kernel is built with -fno-common, so we do instrument
2308 globals there even if it is C. */
2309 || (DECL_COMMON (decl) && TREE_PUBLIC (decl))
2310 /* Don't protect if using user section, often vars placed
2311 into user section from multiple TUs are then assumed
2312 to be an array of such vars, putting padding in there
2313 breaks this assumption. */
2314 || (DECL_SECTION_NAME (decl) != NULL
2315 && !symtab_node::get (decl)->implicit_section
2316 && !section_sanitized_p (DECL_SECTION_NAME (decl)))
2317 /* Don't protect variables in non-generic address-space. */
2318 || !ADDR_SPACE_GENERIC_P (TYPE_ADDR_SPACE (TREE_TYPE (decl)))
2319 || DECL_SIZE (decl) == 0
2320 || ASAN_RED_ZONE_SIZE * BITS_PER_UNIT > MAX_OFILE_ALIGNMENT
2321 || TREE_CODE (DECL_SIZE_UNIT (decl)) != INTEGER_CST
2322 || !valid_constant_size_p (DECL_SIZE_UNIT (decl))
2323 || DECL_ALIGN_UNIT (decl) > 2 * ASAN_RED_ZONE_SIZE
2324 || TREE_TYPE (decl) == ubsan_get_source_location_type ()
2325 || is_odr_indicator (decl))
2326 return false;
2327
2328 if (!ignore_decl_rtl_set_p || DECL_RTL_SET_P (decl))
2329 {
2330
2331 rtl = DECL_RTL (decl);
2332 if (!MEM_P (rtl) || GET_CODE (XEXP (rtl, 0)) != SYMBOL_REF)
2333 return false;
2334 symbol = XEXP (rtl, 0);
2335
2336 if (CONSTANT_POOL_ADDRESS_P (symbol)
2337 || TREE_CONSTANT_POOL_ADDRESS_P (symbol))
2338 return false;
2339 }
2340
2341 if (lookup_attribute (attr_name: "weakref", DECL_ATTRIBUTES (decl)))
2342 return false;
2343
2344 if (!TARGET_SUPPORTS_ALIASES && asan_needs_local_alias (decl))
2345 return false;
2346
2347 return true;
2348}
2349
2350/* Construct a function tree for __asan_report_{load,store}{1,2,4,8,16,_n}.
2351 IS_STORE is either 1 (for a store) or 0 (for a load). */
2352
2353static tree
2354report_error_func (bool is_store, bool recover_p, HOST_WIDE_INT size_in_bytes,
2355 int *nargs)
2356{
2357 gcc_assert (!hwasan_sanitize_p ());
2358
2359 static enum built_in_function report[2][2][6]
2360 = { { { BUILT_IN_ASAN_REPORT_LOAD1, BUILT_IN_ASAN_REPORT_LOAD2,
2361 BUILT_IN_ASAN_REPORT_LOAD4, BUILT_IN_ASAN_REPORT_LOAD8,
2362 BUILT_IN_ASAN_REPORT_LOAD16, BUILT_IN_ASAN_REPORT_LOAD_N },
2363 { BUILT_IN_ASAN_REPORT_STORE1, BUILT_IN_ASAN_REPORT_STORE2,
2364 BUILT_IN_ASAN_REPORT_STORE4, BUILT_IN_ASAN_REPORT_STORE8,
2365 BUILT_IN_ASAN_REPORT_STORE16, BUILT_IN_ASAN_REPORT_STORE_N } },
2366 { { BUILT_IN_ASAN_REPORT_LOAD1_NOABORT,
2367 BUILT_IN_ASAN_REPORT_LOAD2_NOABORT,
2368 BUILT_IN_ASAN_REPORT_LOAD4_NOABORT,
2369 BUILT_IN_ASAN_REPORT_LOAD8_NOABORT,
2370 BUILT_IN_ASAN_REPORT_LOAD16_NOABORT,
2371 BUILT_IN_ASAN_REPORT_LOAD_N_NOABORT },
2372 { BUILT_IN_ASAN_REPORT_STORE1_NOABORT,
2373 BUILT_IN_ASAN_REPORT_STORE2_NOABORT,
2374 BUILT_IN_ASAN_REPORT_STORE4_NOABORT,
2375 BUILT_IN_ASAN_REPORT_STORE8_NOABORT,
2376 BUILT_IN_ASAN_REPORT_STORE16_NOABORT,
2377 BUILT_IN_ASAN_REPORT_STORE_N_NOABORT } } };
2378 if (size_in_bytes == -1)
2379 {
2380 *nargs = 2;
2381 return builtin_decl_implicit (fncode: report[recover_p][is_store][5]);
2382 }
2383 *nargs = 1;
2384 int size_log2 = exact_log2 (x: size_in_bytes);
2385 return builtin_decl_implicit (fncode: report[recover_p][is_store][size_log2]);
2386}
2387
2388/* Construct a function tree for __asan_{load,store}{1,2,4,8,16,_n}.
2389 IS_STORE is either 1 (for a store) or 0 (for a load). */
2390
2391static tree
2392check_func (bool is_store, bool recover_p, HOST_WIDE_INT size_in_bytes,
2393 int *nargs)
2394{
2395 static enum built_in_function check[2][2][6]
2396 = { { { BUILT_IN_ASAN_LOAD1, BUILT_IN_ASAN_LOAD2,
2397 BUILT_IN_ASAN_LOAD4, BUILT_IN_ASAN_LOAD8,
2398 BUILT_IN_ASAN_LOAD16, BUILT_IN_ASAN_LOADN },
2399 { BUILT_IN_ASAN_STORE1, BUILT_IN_ASAN_STORE2,
2400 BUILT_IN_ASAN_STORE4, BUILT_IN_ASAN_STORE8,
2401 BUILT_IN_ASAN_STORE16, BUILT_IN_ASAN_STOREN } },
2402 { { BUILT_IN_ASAN_LOAD1_NOABORT,
2403 BUILT_IN_ASAN_LOAD2_NOABORT,
2404 BUILT_IN_ASAN_LOAD4_NOABORT,
2405 BUILT_IN_ASAN_LOAD8_NOABORT,
2406 BUILT_IN_ASAN_LOAD16_NOABORT,
2407 BUILT_IN_ASAN_LOADN_NOABORT },
2408 { BUILT_IN_ASAN_STORE1_NOABORT,
2409 BUILT_IN_ASAN_STORE2_NOABORT,
2410 BUILT_IN_ASAN_STORE4_NOABORT,
2411 BUILT_IN_ASAN_STORE8_NOABORT,
2412 BUILT_IN_ASAN_STORE16_NOABORT,
2413 BUILT_IN_ASAN_STOREN_NOABORT } } };
2414 if (size_in_bytes == -1)
2415 {
2416 *nargs = 2;
2417 return builtin_decl_implicit (fncode: check[recover_p][is_store][5]);
2418 }
2419 *nargs = 1;
2420 int size_log2 = exact_log2 (x: size_in_bytes);
2421 return builtin_decl_implicit (fncode: check[recover_p][is_store][size_log2]);
2422}
2423
2424/* Split the current basic block and create a condition statement
2425 insertion point right before or after the statement pointed to by
2426 ITER. Return an iterator to the point at which the caller might
2427 safely insert the condition statement.
2428
2429 THEN_BLOCK must be set to the address of an uninitialized instance
2430 of basic_block. The function will then set *THEN_BLOCK to the
2431 'then block' of the condition statement to be inserted by the
2432 caller.
2433
2434 If CREATE_THEN_FALLTHRU_EDGE is false, no edge will be created from
2435 *THEN_BLOCK to *FALLTHROUGH_BLOCK.
2436
2437 Similarly, the function will set *FALLTRHOUGH_BLOCK to the 'else
2438 block' of the condition statement to be inserted by the caller.
2439
2440 Note that *FALLTHROUGH_BLOCK is a new block that contains the
2441 statements starting from *ITER, and *THEN_BLOCK is a new empty
2442 block.
2443
2444 *ITER is adjusted to point to always point to the first statement
2445 of the basic block * FALLTHROUGH_BLOCK. That statement is the
2446 same as what ITER was pointing to prior to calling this function,
2447 if BEFORE_P is true; otherwise, it is its following statement. */
2448
2449gimple_stmt_iterator
2450create_cond_insert_point (gimple_stmt_iterator *iter,
2451 bool before_p,
2452 bool then_more_likely_p,
2453 bool create_then_fallthru_edge,
2454 basic_block *then_block,
2455 basic_block *fallthrough_block)
2456{
2457 gimple_stmt_iterator gsi = *iter;
2458
2459 if (!gsi_end_p (i: gsi) && before_p)
2460 gsi_prev (i: &gsi);
2461
2462 basic_block cur_bb = gsi_bb (i: *iter);
2463
2464 edge e = split_block (cur_bb, gsi_stmt (i: gsi));
2465
2466 /* Get a hold on the 'condition block', the 'then block' and the
2467 'else block'. */
2468 basic_block cond_bb = e->src;
2469 basic_block fallthru_bb = e->dest;
2470 basic_block then_bb = create_empty_bb (cond_bb);
2471 if (current_loops)
2472 {
2473 add_bb_to_loop (then_bb, cond_bb->loop_father);
2474 loops_state_set (flags: LOOPS_NEED_FIXUP);
2475 }
2476
2477 /* Set up the newly created 'then block'. */
2478 e = make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE);
2479 profile_probability fallthrough_probability
2480 = then_more_likely_p
2481 ? profile_probability::very_unlikely ()
2482 : profile_probability::very_likely ();
2483 e->probability = fallthrough_probability.invert ();
2484 then_bb->count = e->count ();
2485 if (create_then_fallthru_edge)
2486 make_single_succ_edge (then_bb, fallthru_bb, EDGE_FALLTHRU);
2487
2488 /* Set up the fallthrough basic block. */
2489 e = find_edge (cond_bb, fallthru_bb);
2490 e->flags = EDGE_FALSE_VALUE;
2491 e->probability = fallthrough_probability;
2492
2493 /* Update dominance info for the newly created then_bb; note that
2494 fallthru_bb's dominance info has already been updated by
2495 split_bock. */
2496 if (dom_info_available_p (CDI_DOMINATORS))
2497 set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb);
2498
2499 *then_block = then_bb;
2500 *fallthrough_block = fallthru_bb;
2501 *iter = gsi_start_bb (bb: fallthru_bb);
2502
2503 return gsi_last_bb (bb: cond_bb);
2504}
2505
2506/* Insert an if condition followed by a 'then block' right before the
2507 statement pointed to by ITER. The fallthrough block -- which is the
2508 else block of the condition as well as the destination of the
2509 outcoming edge of the 'then block' -- starts with the statement
2510 pointed to by ITER.
2511
2512 COND is the condition of the if.
2513
2514 If THEN_MORE_LIKELY_P is true, the probability of the edge to the
2515 'then block' is higher than the probability of the edge to the
2516 fallthrough block.
2517
2518 Upon completion of the function, *THEN_BB is set to the newly
2519 inserted 'then block' and similarly, *FALLTHROUGH_BB is set to the
2520 fallthrough block.
2521
2522 *ITER is adjusted to still point to the same statement it was
2523 pointing to initially. */
2524
2525static void
2526insert_if_then_before_iter (gcond *cond,
2527 gimple_stmt_iterator *iter,
2528 bool then_more_likely_p,
2529 basic_block *then_bb,
2530 basic_block *fallthrough_bb)
2531{
2532 gimple_stmt_iterator cond_insert_point =
2533 create_cond_insert_point (iter,
2534 /*before_p=*/true,
2535 then_more_likely_p,
2536 /*create_then_fallthru_edge=*/true,
2537 then_block: then_bb,
2538 fallthrough_block: fallthrough_bb);
2539 gsi_insert_after (&cond_insert_point, cond, GSI_NEW_STMT);
2540}
2541
2542/* Build (base_addr >> ASAN_SHADOW_SHIFT) + asan_shadow_offset ().
2543 If RETURN_ADDRESS is set to true, return memory location instread
2544 of a value in the shadow memory. */
2545
2546static tree
2547build_shadow_mem_access (gimple_stmt_iterator *gsi, location_t location,
2548 tree base_addr, tree shadow_ptr_type,
2549 bool return_address = false)
2550{
2551 tree t, uintptr_type = TREE_TYPE (base_addr);
2552 tree shadow_type = TREE_TYPE (shadow_ptr_type);
2553 gimple *g;
2554
2555 t = build_int_cst (uintptr_type, ASAN_SHADOW_SHIFT);
2556 g = gimple_build_assign (make_ssa_name (var: uintptr_type), RSHIFT_EXPR,
2557 base_addr, t);
2558 gimple_set_location (g, location);
2559 gsi_insert_after (gsi, g, GSI_NEW_STMT);
2560
2561 t = build_int_cst (uintptr_type, asan_shadow_offset ());
2562 g = gimple_build_assign (make_ssa_name (var: uintptr_type), PLUS_EXPR,
2563 gimple_assign_lhs (gs: g), t);
2564 gimple_set_location (g, location);
2565 gsi_insert_after (gsi, g, GSI_NEW_STMT);
2566
2567 g = gimple_build_assign (make_ssa_name (var: shadow_ptr_type), NOP_EXPR,
2568 gimple_assign_lhs (gs: g));
2569 gimple_set_location (g, location);
2570 gsi_insert_after (gsi, g, GSI_NEW_STMT);
2571
2572 if (!return_address)
2573 {
2574 t = build2 (MEM_REF, shadow_type, gimple_assign_lhs (gs: g),
2575 build_int_cst (shadow_ptr_type, 0));
2576 g = gimple_build_assign (make_ssa_name (var: shadow_type), MEM_REF, t);
2577 gimple_set_location (g, location);
2578 gsi_insert_after (gsi, g, GSI_NEW_STMT);
2579 }
2580
2581 return gimple_assign_lhs (gs: g);
2582}
2583
2584/* BASE can already be an SSA_NAME; in that case, do not create a
2585 new SSA_NAME for it. */
2586
2587static tree
2588maybe_create_ssa_name (location_t loc, tree base, gimple_stmt_iterator *iter,
2589 bool before_p)
2590{
2591 STRIP_USELESS_TYPE_CONVERSION (base);
2592 if (TREE_CODE (base) == SSA_NAME)
2593 return base;
2594 gimple *g = gimple_build_assign (make_ssa_name (TREE_TYPE (base)), base);
2595 gimple_set_location (g, location: loc);
2596 if (before_p)
2597 gsi_safe_insert_before (iter, g);
2598 else
2599 gsi_insert_after (iter, g, GSI_NEW_STMT);
2600 return gimple_assign_lhs (gs: g);
2601}
2602
2603/* LEN can already have necessary size and precision;
2604 in that case, do not create a new variable. */
2605
2606tree
2607maybe_cast_to_ptrmode (location_t loc, tree len, gimple_stmt_iterator *iter,
2608 bool before_p)
2609{
2610 if (ptrofftype_p (type: len))
2611 return len;
2612 gimple *g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
2613 NOP_EXPR, len);
2614 gimple_set_location (g, location: loc);
2615 if (before_p)
2616 gsi_safe_insert_before (iter, g);
2617 else
2618 gsi_insert_after (iter, g, GSI_NEW_STMT);
2619 return gimple_assign_lhs (gs: g);
2620}
2621
2622/* Instrument the memory access instruction BASE. Insert new
2623 statements before or after ITER.
2624
2625 Note that the memory access represented by BASE can be either an
2626 SSA_NAME, or a non-SSA expression. LOCATION is the source code
2627 location. IS_STORE is TRUE for a store, FALSE for a load.
2628 BEFORE_P is TRUE for inserting the instrumentation code before
2629 ITER, FALSE for inserting it after ITER. IS_SCALAR_ACCESS is TRUE
2630 for a scalar memory access and FALSE for memory region access.
2631 NON_ZERO_P is TRUE if memory region is guaranteed to have non-zero
2632 length. ALIGN tells alignment of accessed memory object.
2633
2634 START_INSTRUMENTED and END_INSTRUMENTED are TRUE if start/end of
2635 memory region have already been instrumented.
2636
2637 If BEFORE_P is TRUE, *ITER is arranged to still point to the
2638 statement it was pointing to prior to calling this function,
2639 otherwise, it points to the statement logically following it. */
2640
2641static void
2642build_check_stmt (location_t loc, tree base, tree len,
2643 HOST_WIDE_INT size_in_bytes, gimple_stmt_iterator *iter,
2644 bool is_non_zero_len, bool before_p, bool is_store,
2645 bool is_scalar_access, unsigned int align = 0)
2646{
2647 gimple_stmt_iterator gsi = *iter;
2648 gimple *g;
2649
2650 gcc_assert (!(size_in_bytes > 0 && !is_non_zero_len));
2651 gcc_assert (size_in_bytes == -1 || size_in_bytes >= 1);
2652
2653 gsi = *iter;
2654
2655 base = unshare_expr (base);
2656 base = maybe_create_ssa_name (loc, base, iter: &gsi, before_p);
2657
2658 if (len)
2659 {
2660 len = unshare_expr (len);
2661 len = maybe_cast_to_ptrmode (loc, len, iter, before_p);
2662 }
2663 else
2664 {
2665 gcc_assert (size_in_bytes != -1);
2666 len = build_int_cst (pointer_sized_int_node, size_in_bytes);
2667 }
2668
2669 if (size_in_bytes > 1)
2670 {
2671 if ((size_in_bytes & (size_in_bytes - 1)) != 0
2672 || size_in_bytes > 16)
2673 is_scalar_access = false;
2674 else if (align && align < size_in_bytes * BITS_PER_UNIT)
2675 {
2676 /* On non-strict alignment targets, if
2677 16-byte access is just 8-byte aligned,
2678 this will result in misaligned shadow
2679 memory 2 byte load, but otherwise can
2680 be handled using one read. */
2681 if (size_in_bytes != 16
2682 || STRICT_ALIGNMENT
2683 || align < 8 * BITS_PER_UNIT)
2684 is_scalar_access = false;
2685 }
2686 }
2687
2688 HOST_WIDE_INT flags = 0;
2689 if (is_store)
2690 flags |= ASAN_CHECK_STORE;
2691 if (is_non_zero_len)
2692 flags |= ASAN_CHECK_NON_ZERO_LEN;
2693 if (is_scalar_access)
2694 flags |= ASAN_CHECK_SCALAR_ACCESS;
2695
2696 enum internal_fn fn = hwasan_sanitize_p ()
2697 ? IFN_HWASAN_CHECK
2698 : IFN_ASAN_CHECK;
2699
2700 g = gimple_build_call_internal (fn, 4,
2701 build_int_cst (integer_type_node, flags),
2702 base, len,
2703 build_int_cst (integer_type_node,
2704 align / BITS_PER_UNIT));
2705 gimple_set_location (g, location: loc);
2706 if (before_p)
2707 gsi_safe_insert_before (&gsi, g);
2708 else
2709 {
2710 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
2711 gsi_next (i: &gsi);
2712 *iter = gsi;
2713 }
2714}
2715
2716/* If T represents a memory access, add instrumentation code before ITER.
2717 LOCATION is source code location.
2718 IS_STORE is either TRUE (for a store) or FALSE (for a load). */
2719
2720static void
2721instrument_derefs (gimple_stmt_iterator *iter, tree t,
2722 location_t location, bool is_store)
2723{
2724 if (is_store && !(asan_instrument_writes () || hwasan_instrument_writes ()))
2725 return;
2726 if (!is_store && !(asan_instrument_reads () || hwasan_instrument_reads ()))
2727 return;
2728
2729 tree type, base;
2730 HOST_WIDE_INT size_in_bytes;
2731 if (location == UNKNOWN_LOCATION)
2732 location = EXPR_LOCATION (t);
2733
2734 type = TREE_TYPE (t);
2735 switch (TREE_CODE (t))
2736 {
2737 case ARRAY_REF:
2738 case COMPONENT_REF:
2739 case INDIRECT_REF:
2740 case MEM_REF:
2741 case VAR_DECL:
2742 case BIT_FIELD_REF:
2743 break;
2744 /* FALLTHRU */
2745 default:
2746 return;
2747 }
2748
2749 size_in_bytes = int_size_in_bytes (type);
2750 if (size_in_bytes <= 0)
2751 return;
2752
2753 poly_int64 bitsize, bitpos;
2754 tree offset;
2755 machine_mode mode;
2756 int unsignedp, reversep, volatilep = 0;
2757 tree inner = get_inner_reference (t, &bitsize, &bitpos, &offset, &mode,
2758 &unsignedp, &reversep, &volatilep);
2759
2760 if (TREE_CODE (t) == COMPONENT_REF
2761 && DECL_BIT_FIELD_REPRESENTATIVE (TREE_OPERAND (t, 1)) != NULL_TREE)
2762 {
2763 tree repr = DECL_BIT_FIELD_REPRESENTATIVE (TREE_OPERAND (t, 1));
2764 instrument_derefs (iter, t: build3 (COMPONENT_REF, TREE_TYPE (repr),
2765 TREE_OPERAND (t, 0), repr,
2766 TREE_OPERAND (t, 2)),
2767 location, is_store);
2768 return;
2769 }
2770
2771 if (!multiple_p (a: bitpos, BITS_PER_UNIT)
2772 || maybe_ne (a: bitsize, b: size_in_bytes * BITS_PER_UNIT))
2773 return;
2774
2775 if (VAR_P (inner) && DECL_HARD_REGISTER (inner))
2776 return;
2777
2778 /* Accesses to non-generic address-spaces should not be instrumented. */
2779 if (!ADDR_SPACE_GENERIC_P (TYPE_ADDR_SPACE (TREE_TYPE (inner))))
2780 return;
2781
2782 poly_int64 decl_size;
2783 if ((VAR_P (inner)
2784 || (TREE_CODE (inner) == RESULT_DECL
2785 && !aggregate_value_p (inner, current_function_decl)))
2786 && offset == NULL_TREE
2787 && DECL_SIZE (inner)
2788 && poly_int_tree_p (DECL_SIZE (inner), value: &decl_size)
2789 && known_subrange_p (pos1: bitpos, size1: bitsize, pos2: 0, size2: decl_size))
2790 {
2791 if (VAR_P (inner) && DECL_THREAD_LOCAL_P (inner))
2792 return;
2793 /* If we're not sanitizing globals and we can tell statically that this
2794 access is inside a global variable, then there's no point adding
2795 instrumentation to check the access. N.b. hwasan currently never
2796 sanitizes globals. */
2797 if ((hwasan_sanitize_p () || !param_asan_globals)
2798 && is_global_var (t: inner))
2799 return;
2800 if (!TREE_STATIC (inner))
2801 {
2802 /* Automatic vars in the current function will be always
2803 accessible. */
2804 if (decl_function_context (inner) == current_function_decl
2805 && (!asan_sanitize_use_after_scope ()
2806 || !TREE_ADDRESSABLE (inner)))
2807 return;
2808 }
2809 /* Always instrument external vars, they might be dynamically
2810 initialized. */
2811 else if (!DECL_EXTERNAL (inner))
2812 {
2813 /* For static vars if they are known not to be dynamically
2814 initialized, they will be always accessible. */
2815 varpool_node *vnode = varpool_node::get (decl: inner);
2816 if (vnode && !vnode->dynamically_initialized)
2817 return;
2818 }
2819 }
2820
2821 if (DECL_P (inner)
2822 && decl_function_context (inner) == current_function_decl
2823 && !TREE_ADDRESSABLE (inner))
2824 mark_addressable (inner);
2825
2826 base = build_fold_addr_expr (t);
2827 if (!has_mem_ref_been_instrumented (ref: base, access_size: size_in_bytes))
2828 {
2829 unsigned int align = get_object_alignment (t);
2830 build_check_stmt (loc: location, base, NULL_TREE, size_in_bytes, iter,
2831 /*is_non_zero_len*/size_in_bytes > 0, /*before_p=*/true,
2832 is_store, /*is_scalar_access*/true, align);
2833 update_mem_ref_hash_table (ref: base, access_size: size_in_bytes);
2834 update_mem_ref_hash_table (ref: t, access_size: size_in_bytes);
2835 }
2836
2837}
2838
2839/* Insert a memory reference into the hash table if access length
2840 can be determined in compile time. */
2841
2842static void
2843maybe_update_mem_ref_hash_table (tree base, tree len)
2844{
2845 if (!POINTER_TYPE_P (TREE_TYPE (base))
2846 || !INTEGRAL_TYPE_P (TREE_TYPE (len)))
2847 return;
2848
2849 HOST_WIDE_INT size_in_bytes = tree_fits_shwi_p (len) ? tree_to_shwi (len) : -1;
2850
2851 if (size_in_bytes != -1)
2852 update_mem_ref_hash_table (ref: base, access_size: size_in_bytes);
2853}
2854
2855/* Instrument an access to a contiguous memory region that starts at
2856 the address pointed to by BASE, over a length of LEN (expressed in
2857 the sizeof (*BASE) bytes). ITER points to the instruction before
2858 which the instrumentation instructions must be inserted. LOCATION
2859 is the source location that the instrumentation instructions must
2860 have. If IS_STORE is true, then the memory access is a store;
2861 otherwise, it's a load. */
2862
2863static void
2864instrument_mem_region_access (tree base, tree len,
2865 gimple_stmt_iterator *iter,
2866 location_t location, bool is_store)
2867{
2868 if (!POINTER_TYPE_P (TREE_TYPE (base))
2869 || !INTEGRAL_TYPE_P (TREE_TYPE (len))
2870 || integer_zerop (len))
2871 return;
2872
2873 HOST_WIDE_INT size_in_bytes = tree_fits_shwi_p (len) ? tree_to_shwi (len) : -1;
2874
2875 if ((size_in_bytes == -1)
2876 || !has_mem_ref_been_instrumented (ref: base, access_size: size_in_bytes))
2877 {
2878 build_check_stmt (loc: location, base, len, size_in_bytes, iter,
2879 /*is_non_zero_len*/size_in_bytes > 0, /*before_p*/true,
2880 is_store, /*is_scalar_access*/false, /*align*/0);
2881 }
2882
2883 maybe_update_mem_ref_hash_table (base, len);
2884 *iter = gsi_for_stmt (gsi_stmt (i: *iter));
2885}
2886
2887/* Instrument the call to a built-in memory access function that is
2888 pointed to by the iterator ITER.
2889
2890 Upon completion, return TRUE iff *ITER has been advanced to the
2891 statement following the one it was originally pointing to. */
2892
2893static bool
2894instrument_builtin_call (gimple_stmt_iterator *iter)
2895{
2896 if (!(asan_memintrin () || hwasan_memintrin ()))
2897 return false;
2898
2899 bool iter_advanced_p = false;
2900 gcall *call = as_a <gcall *> (p: gsi_stmt (i: *iter));
2901
2902 gcc_checking_assert (gimple_call_builtin_p (call, BUILT_IN_NORMAL));
2903
2904 location_t loc = gimple_location (g: call);
2905
2906 asan_mem_ref src0, src1, dest;
2907 asan_mem_ref_init (ref: &src0, NULL, access_size: 1);
2908 asan_mem_ref_init (ref: &src1, NULL, access_size: 1);
2909 asan_mem_ref_init (ref: &dest, NULL, access_size: 1);
2910
2911 tree src0_len = NULL_TREE, src1_len = NULL_TREE, dest_len = NULL_TREE;
2912 bool src0_is_store = false, src1_is_store = false, dest_is_store = false,
2913 dest_is_deref = false, intercepted_p = true;
2914
2915 if (get_mem_refs_of_builtin_call (call,
2916 src0: &src0, src0_len: &src0_len, src0_is_store: &src0_is_store,
2917 src1: &src1, src1_len: &src1_len, src1_is_store: &src1_is_store,
2918 dst: &dest, dst_len: &dest_len, dst_is_store: &dest_is_store,
2919 dest_is_deref: &dest_is_deref, intercepted_p: &intercepted_p, iter))
2920 {
2921 if (dest_is_deref)
2922 {
2923 instrument_derefs (iter, t: dest.start, location: loc, is_store: dest_is_store);
2924 gsi_next (i: iter);
2925 iter_advanced_p = true;
2926 }
2927 else if (!intercepted_p
2928 && (src0_len || src1_len || dest_len))
2929 {
2930 if (src0.start != NULL_TREE)
2931 instrument_mem_region_access (base: src0.start, len: src0_len,
2932 iter, location: loc, /*is_store=*/false);
2933 if (src1.start != NULL_TREE)
2934 instrument_mem_region_access (base: src1.start, len: src1_len,
2935 iter, location: loc, /*is_store=*/false);
2936 if (dest.start != NULL_TREE)
2937 instrument_mem_region_access (base: dest.start, len: dest_len,
2938 iter, location: loc, /*is_store=*/true);
2939
2940 *iter = gsi_for_stmt (call);
2941 gsi_next (i: iter);
2942 iter_advanced_p = true;
2943 }
2944 else
2945 {
2946 if (src0.start != NULL_TREE)
2947 maybe_update_mem_ref_hash_table (base: src0.start, len: src0_len);
2948 if (src1.start != NULL_TREE)
2949 maybe_update_mem_ref_hash_table (base: src1.start, len: src1_len);
2950 if (dest.start != NULL_TREE)
2951 maybe_update_mem_ref_hash_table (base: dest.start, len: dest_len);
2952 }
2953 }
2954 return iter_advanced_p;
2955}
2956
2957/* Instrument the assignment statement ITER if it is subject to
2958 instrumentation. Return TRUE iff instrumentation actually
2959 happened. In that case, the iterator ITER is advanced to the next
2960 logical expression following the one initially pointed to by ITER,
2961 and the relevant memory reference that which access has been
2962 instrumented is added to the memory references hash table. */
2963
2964static bool
2965maybe_instrument_assignment (gimple_stmt_iterator *iter)
2966{
2967 gimple *s = gsi_stmt (i: *iter);
2968
2969 gcc_assert (gimple_assign_single_p (s));
2970
2971 tree ref_expr = NULL_TREE;
2972 bool is_store, is_instrumented = false;
2973
2974 if (gimple_store_p (gs: s))
2975 {
2976 ref_expr = gimple_assign_lhs (gs: s);
2977 is_store = true;
2978 instrument_derefs (iter, t: ref_expr,
2979 location: gimple_location (g: s),
2980 is_store);
2981 is_instrumented = true;
2982 }
2983
2984 if (gimple_assign_load_p (s))
2985 {
2986 ref_expr = gimple_assign_rhs1 (gs: s);
2987 is_store = false;
2988 instrument_derefs (iter, t: ref_expr,
2989 location: gimple_location (g: s),
2990 is_store);
2991 is_instrumented = true;
2992 }
2993
2994 if (is_instrumented)
2995 gsi_next (i: iter);
2996
2997 return is_instrumented;
2998}
2999
3000/* Instrument the function call pointed to by the iterator ITER, if it
3001 is subject to instrumentation. At the moment, the only function
3002 calls that are instrumented are some built-in functions that access
3003 memory. Look at instrument_builtin_call to learn more.
3004
3005 Upon completion return TRUE iff *ITER was advanced to the statement
3006 following the one it was originally pointing to. */
3007
3008static bool
3009maybe_instrument_call (gimple_stmt_iterator *iter)
3010{
3011 gimple *stmt = gsi_stmt (i: *iter);
3012 bool is_builtin = gimple_call_builtin_p (stmt, BUILT_IN_NORMAL);
3013
3014 if (is_builtin && instrument_builtin_call (iter))
3015 return true;
3016
3017 if (gimple_call_noreturn_p (s: stmt))
3018 {
3019 if (is_builtin)
3020 {
3021 tree callee = gimple_call_fndecl (gs: stmt);
3022 switch (DECL_FUNCTION_CODE (decl: callee))
3023 {
3024 case BUILT_IN_UNREACHABLE:
3025 case BUILT_IN_UNREACHABLE_TRAP:
3026 case BUILT_IN_TRAP:
3027 /* Don't instrument these. */
3028 return false;
3029 default:
3030 break;
3031 }
3032 }
3033 if (gimple_call_internal_p (gs: stmt, fn: IFN_ABNORMAL_DISPATCHER))
3034 /* Don't instrument this. */
3035 return false;
3036 /* If a function does not return, then we must handle clearing up the
3037 shadow stack accordingly. For ASAN we can simply set the entire stack
3038 to "valid" for accesses by setting the shadow space to 0 and all
3039 accesses will pass checks. That means that some bad accesses may be
3040 missed, but we will not report any false positives.
3041
3042 This is not possible for HWASAN. Since there is no "always valid" tag
3043 we can not set any space to "always valid". If we were to clear the
3044 entire shadow stack then code resuming from `longjmp` or a caught
3045 exception would trigger false positives when correctly accessing
3046 variables on the stack. Hence we need to handle things like
3047 `longjmp`, thread exit, and exceptions in a different way. These
3048 problems must be handled externally to the compiler, e.g. in the
3049 language runtime. */
3050 if (! hwasan_sanitize_p ())
3051 {
3052 tree decl = builtin_decl_implicit (fncode: BUILT_IN_ASAN_HANDLE_NO_RETURN);
3053 gimple *g = gimple_build_call (decl, 0);
3054 gimple_set_location (g, location: gimple_location (g: stmt));
3055 gsi_safe_insert_before (iter, g);
3056 }
3057 }
3058
3059 bool instrumented = false;
3060 if (gimple_store_p (gs: stmt)
3061 && (gimple_call_builtin_p (stmt)
3062 || gimple_call_internal_p (gs: stmt)
3063 || !aggregate_value_p (TREE_TYPE (gimple_call_lhs (stmt)),
3064 gimple_call_fntype (gs: stmt))))
3065 {
3066 tree ref_expr = gimple_call_lhs (gs: stmt);
3067 instrument_derefs (iter, t: ref_expr,
3068 location: gimple_location (g: stmt),
3069 /*is_store=*/true);
3070
3071 instrumented = true;
3072 }
3073
3074 /* Walk through gimple_call arguments and check them id needed. */
3075 unsigned args_num = gimple_call_num_args (gs: stmt);
3076 for (unsigned i = 0; i < args_num; ++i)
3077 {
3078 tree arg = gimple_call_arg (gs: stmt, index: i);
3079 /* If ARG is not a non-aggregate register variable, compiler in general
3080 creates temporary for it and pass it as argument to gimple call.
3081 But in some cases, e.g. when we pass by value a small structure that
3082 fits to register, compiler can avoid extra overhead by pulling out
3083 these temporaries. In this case, we should check the argument. */
3084 if (!is_gimple_reg (arg) && !is_gimple_min_invariant (arg))
3085 {
3086 instrument_derefs (iter, t: arg,
3087 location: gimple_location (g: stmt),
3088 /*is_store=*/false);
3089 instrumented = true;
3090 }
3091 }
3092 if (instrumented)
3093 gsi_next (i: iter);
3094 return instrumented;
3095}
3096
3097/* Walk each instruction of all basic block and instrument those that
3098 represent memory references: loads, stores, or function calls.
3099 In a given basic block, this function avoids instrumenting memory
3100 references that have already been instrumented. */
3101
3102static void
3103transform_statements (void)
3104{
3105 basic_block bb, last_bb = NULL;
3106 gimple_stmt_iterator i;
3107 int saved_last_basic_block = last_basic_block_for_fn (cfun);
3108
3109 FOR_EACH_BB_FN (bb, cfun)
3110 {
3111 basic_block prev_bb = bb;
3112
3113 if (bb->index >= saved_last_basic_block) continue;
3114
3115 /* Flush the mem ref hash table, if current bb doesn't have
3116 exactly one predecessor, or if that predecessor (skipping
3117 over asan created basic blocks) isn't the last processed
3118 basic block. Thus we effectively flush on extended basic
3119 block boundaries. */
3120 while (single_pred_p (bb: prev_bb))
3121 {
3122 prev_bb = single_pred (bb: prev_bb);
3123 if (prev_bb->index < saved_last_basic_block)
3124 break;
3125 }
3126 if (prev_bb != last_bb)
3127 empty_mem_ref_hash_table ();
3128 last_bb = bb;
3129
3130 for (i = gsi_start_bb (bb); !gsi_end_p (i);)
3131 {
3132 gimple *s = gsi_stmt (i);
3133
3134 if (has_stmt_been_instrumented_p (stmt: s))
3135 gsi_next (i: &i);
3136 else if (gimple_assign_single_p (gs: s)
3137 && !gimple_clobber_p (s)
3138 && maybe_instrument_assignment (iter: &i))
3139 /* Nothing to do as maybe_instrument_assignment advanced
3140 the iterator I. */;
3141 else if (is_gimple_call (gs: s) && maybe_instrument_call (iter: &i))
3142 /* Nothing to do as maybe_instrument_call
3143 advanced the iterator I. */;
3144 else
3145 {
3146 /* No instrumentation happened.
3147
3148 If the current instruction is a function call that
3149 might free something, let's forget about the memory
3150 references that got instrumented. Otherwise we might
3151 miss some instrumentation opportunities. Do the same
3152 for a ASAN_MARK poisoning internal function. */
3153 if (is_gimple_call (gs: s)
3154 && (!nonfreeing_call_p (s)
3155 || asan_mark_p (stmt: s, flag: ASAN_MARK_POISON)))
3156 empty_mem_ref_hash_table ();
3157
3158 gsi_next (i: &i);
3159 }
3160 }
3161 }
3162 free_mem_ref_resources ();
3163}
3164
3165/* Build
3166 __asan_before_dynamic_init (module_name)
3167 or
3168 __asan_after_dynamic_init ()
3169 call. */
3170
3171tree
3172asan_dynamic_init_call (bool after_p)
3173{
3174 if (shadow_ptr_types[0] == NULL_TREE)
3175 asan_init_shadow_ptr_types ();
3176
3177 tree fn = builtin_decl_implicit (fncode: after_p
3178 ? BUILT_IN_ASAN_AFTER_DYNAMIC_INIT
3179 : BUILT_IN_ASAN_BEFORE_DYNAMIC_INIT);
3180 tree module_name_cst = NULL_TREE;
3181 if (!after_p)
3182 {
3183 pretty_printer module_name_pp;
3184 pp_string (&module_name_pp, main_input_filename);
3185
3186 module_name_cst = asan_pp_string (pp: &module_name_pp);
3187 module_name_cst = fold_convert (const_ptr_type_node,
3188 module_name_cst);
3189 }
3190
3191 return build_call_expr (fn, after_p ? 0 : 1, module_name_cst);
3192}
3193
3194/* Build
3195 struct __asan_global
3196 {
3197 const void *__beg;
3198 uptr __size;
3199 uptr __size_with_redzone;
3200 const void *__name;
3201 const void *__module_name;
3202 uptr __has_dynamic_init;
3203 __asan_global_source_location *__location;
3204 char *__odr_indicator;
3205 } type. */
3206
3207static tree
3208asan_global_struct (void)
3209{
3210 static const char *field_names[]
3211 = { "__beg", "__size", "__size_with_redzone",
3212 "__name", "__module_name", "__has_dynamic_init", "__location",
3213 "__odr_indicator" };
3214 tree fields[ARRAY_SIZE (field_names)], ret;
3215 unsigned i;
3216
3217 ret = make_node (RECORD_TYPE);
3218 for (i = 0; i < ARRAY_SIZE (field_names); i++)
3219 {
3220 fields[i]
3221 = build_decl (UNKNOWN_LOCATION, FIELD_DECL,
3222 get_identifier (field_names[i]),
3223 (i == 0 || i == 3) ? const_ptr_type_node
3224 : pointer_sized_int_node);
3225 DECL_CONTEXT (fields[i]) = ret;
3226 if (i)
3227 DECL_CHAIN (fields[i - 1]) = fields[i];
3228 }
3229 tree type_decl = build_decl (input_location, TYPE_DECL,
3230 get_identifier ("__asan_global"), ret);
3231 DECL_IGNORED_P (type_decl) = 1;
3232 DECL_ARTIFICIAL (type_decl) = 1;
3233 TYPE_FIELDS (ret) = fields[0];
3234 TYPE_NAME (ret) = type_decl;
3235 TYPE_STUB_DECL (ret) = type_decl;
3236 TYPE_ARTIFICIAL (ret) = 1;
3237 layout_type (ret);
3238 return ret;
3239}
3240
3241/* Create and return odr indicator symbol for DECL.
3242 TYPE is __asan_global struct type as returned by asan_global_struct. */
3243
3244static tree
3245create_odr_indicator (tree decl, tree type)
3246{
3247 char *name;
3248 tree uptr = TREE_TYPE (DECL_CHAIN (TYPE_FIELDS (type)));
3249 tree decl_name
3250 = (HAS_DECL_ASSEMBLER_NAME_P (decl) ? DECL_ASSEMBLER_NAME (decl)
3251 : DECL_NAME (decl));
3252 /* DECL_NAME theoretically might be NULL. Bail out with 0 in this case. */
3253 if (decl_name == NULL_TREE)
3254 return build_int_cst (uptr, 0);
3255 const char *dname = IDENTIFIER_POINTER (decl_name);
3256 if (HAS_DECL_ASSEMBLER_NAME_P (decl))
3257 dname = targetm.strip_name_encoding (dname);
3258 size_t len = strlen (s: dname) + sizeof ("__odr_asan_");
3259 name = XALLOCAVEC (char, len);
3260 snprintf (s: name, maxlen: len, format: "__odr_asan_%s", dname);
3261#ifndef NO_DOT_IN_LABEL
3262 name[sizeof ("__odr_asan") - 1] = '.';
3263#elif !defined(NO_DOLLAR_IN_LABEL)
3264 name[sizeof ("__odr_asan") - 1] = '$';
3265#endif
3266 tree var = build_decl (UNKNOWN_LOCATION, VAR_DECL, get_identifier (name),
3267 char_type_node);
3268 TREE_ADDRESSABLE (var) = 1;
3269 TREE_READONLY (var) = 0;
3270 TREE_THIS_VOLATILE (var) = 1;
3271 DECL_ARTIFICIAL (var) = 1;
3272 DECL_IGNORED_P (var) = 1;
3273 TREE_STATIC (var) = 1;
3274 TREE_PUBLIC (var) = 1;
3275 DECL_VISIBILITY (var) = DECL_VISIBILITY (decl);
3276 DECL_VISIBILITY_SPECIFIED (var) = DECL_VISIBILITY_SPECIFIED (decl);
3277
3278 TREE_USED (var) = 1;
3279 tree ctor = build_constructor_va (TREE_TYPE (var), 1, NULL_TREE,
3280 build_int_cst (unsigned_type_node, 0));
3281 TREE_CONSTANT (ctor) = 1;
3282 TREE_STATIC (ctor) = 1;
3283 DECL_INITIAL (var) = ctor;
3284 DECL_ATTRIBUTES (var) = tree_cons (get_identifier ("asan odr indicator"),
3285 NULL, DECL_ATTRIBUTES (var));
3286 make_decl_rtl (var);
3287 varpool_node::finalize_decl (decl: var);
3288 return fold_convert (uptr, build_fold_addr_expr (var));
3289}
3290
3291/* Return true if DECL, a global var, might be overridden and needs
3292 an additional odr indicator symbol. */
3293
3294static bool
3295asan_needs_odr_indicator_p (tree decl)
3296{
3297 /* Don't emit ODR indicators for kernel because:
3298 a) Kernel is written in C thus doesn't need ODR indicators.
3299 b) Some kernel code may have assumptions about symbols containing specific
3300 patterns in their names. Since ODR indicators contain original names
3301 of symbols they are emitted for, these assumptions would be broken for
3302 ODR indicator symbols. */
3303 return (!(flag_sanitize & SANITIZE_KERNEL_ADDRESS)
3304 && !DECL_ARTIFICIAL (decl)
3305 && !DECL_WEAK (decl)
3306 && TREE_PUBLIC (decl));
3307}
3308
3309/* Append description of a single global DECL into vector V.
3310 TYPE is __asan_global struct type as returned by asan_global_struct. */
3311
3312static void
3313asan_add_global (tree decl, tree type, vec<constructor_elt, va_gc> *v)
3314{
3315 tree init, uptr = TREE_TYPE (DECL_CHAIN (TYPE_FIELDS (type)));
3316 unsigned HOST_WIDE_INT size;
3317 tree str_cst, module_name_cst, refdecl = decl;
3318 vec<constructor_elt, va_gc> *vinner = NULL;
3319
3320 pretty_printer asan_pp, module_name_pp;
3321
3322 if (DECL_NAME (decl))
3323 pp_tree_identifier (&asan_pp, DECL_NAME (decl));
3324 else
3325 pp_string (&asan_pp, "<unknown>");
3326 str_cst = asan_pp_string (pp: &asan_pp);
3327
3328 if (!in_lto_p)
3329 pp_string (&module_name_pp, main_input_filename);
3330 else
3331 {
3332 const_tree tu = get_ultimate_context ((const_tree)decl);
3333 if (tu != NULL_TREE)
3334 pp_string (&module_name_pp, IDENTIFIER_POINTER (DECL_NAME (tu)));
3335 else
3336 pp_string (&module_name_pp, aux_base_name);
3337 }
3338
3339 module_name_cst = asan_pp_string (pp: &module_name_pp);
3340
3341 if (asan_needs_local_alias (decl))
3342 {
3343 char buf[20];
3344 ASM_GENERATE_INTERNAL_LABEL (buf, "LASAN", vec_safe_length (v) + 1);
3345 refdecl = build_decl (DECL_SOURCE_LOCATION (decl),
3346 VAR_DECL, get_identifier (buf), TREE_TYPE (decl));
3347 TREE_ADDRESSABLE (refdecl) = TREE_ADDRESSABLE (decl);
3348 TREE_READONLY (refdecl) = TREE_READONLY (decl);
3349 TREE_THIS_VOLATILE (refdecl) = TREE_THIS_VOLATILE (decl);
3350 DECL_NOT_GIMPLE_REG_P (refdecl) = DECL_NOT_GIMPLE_REG_P (decl);
3351 DECL_ARTIFICIAL (refdecl) = DECL_ARTIFICIAL (decl);
3352 DECL_IGNORED_P (refdecl) = DECL_IGNORED_P (decl);
3353 TREE_STATIC (refdecl) = 1;
3354 TREE_PUBLIC (refdecl) = 0;
3355 TREE_USED (refdecl) = 1;
3356 assemble_alias (refdecl, DECL_ASSEMBLER_NAME (decl));
3357 }
3358
3359 tree odr_indicator_ptr
3360 = (asan_needs_odr_indicator_p (decl) ? create_odr_indicator (decl, type)
3361 : build_int_cst (uptr, 0));
3362 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE,
3363 fold_convert (const_ptr_type_node,
3364 build_fold_addr_expr (refdecl)));
3365 size = tree_to_uhwi (DECL_SIZE_UNIT (decl));
3366 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE, build_int_cst (uptr, size));
3367 size += asan_red_zone_size (size);
3368 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE, build_int_cst (uptr, size));
3369 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE,
3370 fold_convert (const_ptr_type_node, str_cst));
3371 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE,
3372 fold_convert (const_ptr_type_node, module_name_cst));
3373 varpool_node *vnode = varpool_node::get (decl);
3374 int has_dynamic_init = 0;
3375 /* FIXME: Enable initialization order fiasco detection in LTO mode once
3376 proper fix for PR 79061 will be applied. */
3377 if (!in_lto_p)
3378 has_dynamic_init = vnode ? vnode->dynamically_initialized : 0;
3379 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE,
3380 build_int_cst (uptr, has_dynamic_init));
3381 tree locptr = NULL_TREE;
3382 location_t loc = DECL_SOURCE_LOCATION (decl);
3383 expanded_location xloc = expand_location (loc);
3384 if (xloc.file != NULL)
3385 {
3386 static int lasanloccnt = 0;
3387 char buf[25];
3388 ASM_GENERATE_INTERNAL_LABEL (buf, "LASANLOC", ++lasanloccnt);
3389 tree var = build_decl (UNKNOWN_LOCATION, VAR_DECL, get_identifier (buf),
3390 ubsan_get_source_location_type ());
3391 TREE_STATIC (var) = 1;
3392 TREE_PUBLIC (var) = 0;
3393 DECL_ARTIFICIAL (var) = 1;
3394 DECL_IGNORED_P (var) = 1;
3395 pretty_printer filename_pp;
3396 pp_string (&filename_pp, xloc.file);
3397 tree str = asan_pp_string (pp: &filename_pp);
3398 tree ctor = build_constructor_va (TREE_TYPE (var), 3,
3399 NULL_TREE, str, NULL_TREE,
3400 build_int_cst (unsigned_type_node,
3401 xloc.line), NULL_TREE,
3402 build_int_cst (unsigned_type_node,
3403 xloc.column));
3404 TREE_CONSTANT (ctor) = 1;
3405 TREE_STATIC (ctor) = 1;
3406 DECL_INITIAL (var) = ctor;
3407 varpool_node::finalize_decl (decl: var);
3408 locptr = fold_convert (uptr, build_fold_addr_expr (var));
3409 }
3410 else
3411 locptr = build_int_cst (uptr, 0);
3412 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE, locptr);
3413 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE, odr_indicator_ptr);
3414 init = build_constructor (type, vinner);
3415 CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, init);
3416}
3417
3418/* Initialize sanitizer.def builtins if the FE hasn't initialized them. */
3419void
3420initialize_sanitizer_builtins (void)
3421{
3422 tree decl;
3423
3424 if (builtin_decl_implicit_p (fncode: BUILT_IN_ASAN_INIT))
3425 return;
3426
3427 tree BT_FN_VOID = build_function_type_list (void_type_node, NULL_TREE);
3428 tree BT_FN_VOID_PTR
3429 = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE);
3430 tree BT_FN_VOID_CONST_PTR
3431 = build_function_type_list (void_type_node, const_ptr_type_node, NULL_TREE);
3432 tree BT_FN_VOID_PTR_PTR
3433 = build_function_type_list (void_type_node, ptr_type_node,
3434 ptr_type_node, NULL_TREE);
3435 tree BT_FN_VOID_PTR_PTR_PTR
3436 = build_function_type_list (void_type_node, ptr_type_node,
3437 ptr_type_node, ptr_type_node, NULL_TREE);
3438 tree BT_FN_VOID_PTR_PTRMODE
3439 = build_function_type_list (void_type_node, ptr_type_node,
3440 pointer_sized_int_node, NULL_TREE);
3441 tree BT_FN_VOID_INT
3442 = build_function_type_list (void_type_node, integer_type_node, NULL_TREE);
3443 tree BT_FN_SIZE_CONST_PTR_INT
3444 = build_function_type_list (size_type_node, const_ptr_type_node,
3445 integer_type_node, NULL_TREE);
3446
3447 tree BT_FN_VOID_UINT8_UINT8
3448 = build_function_type_list (void_type_node, unsigned_char_type_node,
3449 unsigned_char_type_node, NULL_TREE);
3450 tree BT_FN_VOID_UINT16_UINT16
3451 = build_function_type_list (void_type_node, uint16_type_node,
3452 uint16_type_node, NULL_TREE);
3453 tree BT_FN_VOID_UINT32_UINT32
3454 = build_function_type_list (void_type_node, uint32_type_node,
3455 uint32_type_node, NULL_TREE);
3456 tree BT_FN_VOID_UINT64_UINT64
3457 = build_function_type_list (void_type_node, uint64_type_node,
3458 uint64_type_node, NULL_TREE);
3459 tree BT_FN_VOID_FLOAT_FLOAT
3460 = build_function_type_list (void_type_node, float_type_node,
3461 float_type_node, NULL_TREE);
3462 tree BT_FN_VOID_DOUBLE_DOUBLE
3463 = build_function_type_list (void_type_node, double_type_node,
3464 double_type_node, NULL_TREE);
3465 tree BT_FN_VOID_UINT64_PTR
3466 = build_function_type_list (void_type_node, uint64_type_node,
3467 ptr_type_node, NULL_TREE);
3468
3469 tree BT_FN_PTR_CONST_PTR_UINT8
3470 = build_function_type_list (ptr_type_node, const_ptr_type_node,
3471 unsigned_char_type_node, NULL_TREE);
3472 tree BT_FN_VOID_PTR_UINT8_PTRMODE
3473 = build_function_type_list (void_type_node, ptr_type_node,
3474 unsigned_char_type_node,
3475 pointer_sized_int_node, NULL_TREE);
3476
3477 tree BT_FN_BOOL_VPTR_PTR_IX_INT_INT[5];
3478 tree BT_FN_IX_CONST_VPTR_INT[5];
3479 tree BT_FN_IX_VPTR_IX_INT[5];
3480 tree BT_FN_VOID_VPTR_IX_INT[5];
3481 tree vptr
3482 = build_pointer_type (build_qualified_type (void_type_node,
3483 TYPE_QUAL_VOLATILE));
3484 tree cvptr
3485 = build_pointer_type (build_qualified_type (void_type_node,
3486 TYPE_QUAL_VOLATILE
3487 |TYPE_QUAL_CONST));
3488 tree boolt
3489 = lang_hooks.types.type_for_size (BOOL_TYPE_SIZE, 1);
3490 int i;
3491 for (i = 0; i < 5; i++)
3492 {
3493 tree ix = build_nonstandard_integer_type (BITS_PER_UNIT * (1 << i), 1);
3494 BT_FN_BOOL_VPTR_PTR_IX_INT_INT[i]
3495 = build_function_type_list (boolt, vptr, ptr_type_node, ix,
3496 integer_type_node, integer_type_node,
3497 NULL_TREE);
3498 BT_FN_IX_CONST_VPTR_INT[i]
3499 = build_function_type_list (ix, cvptr, integer_type_node, NULL_TREE);
3500 BT_FN_IX_VPTR_IX_INT[i]
3501 = build_function_type_list (ix, vptr, ix, integer_type_node,
3502 NULL_TREE);
3503 BT_FN_VOID_VPTR_IX_INT[i]
3504 = build_function_type_list (void_type_node, vptr, ix,
3505 integer_type_node, NULL_TREE);
3506 }
3507#define BT_FN_BOOL_VPTR_PTR_I1_INT_INT BT_FN_BOOL_VPTR_PTR_IX_INT_INT[0]
3508#define BT_FN_I1_CONST_VPTR_INT BT_FN_IX_CONST_VPTR_INT[0]
3509#define BT_FN_I1_VPTR_I1_INT BT_FN_IX_VPTR_IX_INT[0]
3510#define BT_FN_VOID_VPTR_I1_INT BT_FN_VOID_VPTR_IX_INT[0]
3511#define BT_FN_BOOL_VPTR_PTR_I2_INT_INT BT_FN_BOOL_VPTR_PTR_IX_INT_INT[1]
3512#define BT_FN_I2_CONST_VPTR_INT BT_FN_IX_CONST_VPTR_INT[1]
3513#define BT_FN_I2_VPTR_I2_INT BT_FN_IX_VPTR_IX_INT[1]
3514#define BT_FN_VOID_VPTR_I2_INT BT_FN_VOID_VPTR_IX_INT[1]
3515#define BT_FN_BOOL_VPTR_PTR_I4_INT_INT BT_FN_BOOL_VPTR_PTR_IX_INT_INT[2]
3516#define BT_FN_I4_CONST_VPTR_INT BT_FN_IX_CONST_VPTR_INT[2]
3517#define BT_FN_I4_VPTR_I4_INT BT_FN_IX_VPTR_IX_INT[2]
3518#define BT_FN_VOID_VPTR_I4_INT BT_FN_VOID_VPTR_IX_INT[2]
3519#define BT_FN_BOOL_VPTR_PTR_I8_INT_INT BT_FN_BOOL_VPTR_PTR_IX_INT_INT[3]
3520#define BT_FN_I8_CONST_VPTR_INT BT_FN_IX_CONST_VPTR_INT[3]
3521#define BT_FN_I8_VPTR_I8_INT BT_FN_IX_VPTR_IX_INT[3]
3522#define BT_FN_VOID_VPTR_I8_INT BT_FN_VOID_VPTR_IX_INT[3]
3523#define BT_FN_BOOL_VPTR_PTR_I16_INT_INT BT_FN_BOOL_VPTR_PTR_IX_INT_INT[4]
3524#define BT_FN_I16_CONST_VPTR_INT BT_FN_IX_CONST_VPTR_INT[4]
3525#define BT_FN_I16_VPTR_I16_INT BT_FN_IX_VPTR_IX_INT[4]
3526#define BT_FN_VOID_VPTR_I16_INT BT_FN_VOID_VPTR_IX_INT[4]
3527#undef ATTR_NOTHROW_LIST
3528#define ATTR_NOTHROW_LIST ECF_NOTHROW
3529#undef ATTR_NOTHROW_LEAF_LIST
3530#define ATTR_NOTHROW_LEAF_LIST ECF_NOTHROW | ECF_LEAF
3531#undef ATTR_TMPURE_NOTHROW_LEAF_LIST
3532#define ATTR_TMPURE_NOTHROW_LEAF_LIST ECF_TM_PURE | ATTR_NOTHROW_LEAF_LIST
3533#undef ATTR_NORETURN_NOTHROW_LEAF_LIST
3534#define ATTR_NORETURN_NOTHROW_LEAF_LIST ECF_NORETURN | ATTR_NOTHROW_LEAF_LIST
3535#undef ATTR_CONST_NORETURN_NOTHROW_LEAF_LIST
3536#define ATTR_CONST_NORETURN_NOTHROW_LEAF_LIST \
3537 ECF_CONST | ATTR_NORETURN_NOTHROW_LEAF_LIST
3538#undef ATTR_TMPURE_NORETURN_NOTHROW_LEAF_LIST
3539#define ATTR_TMPURE_NORETURN_NOTHROW_LEAF_LIST \
3540 ECF_TM_PURE | ATTR_NORETURN_NOTHROW_LEAF_LIST
3541#undef ATTR_COLD_NOTHROW_LEAF_LIST
3542#define ATTR_COLD_NOTHROW_LEAF_LIST \
3543 /* ECF_COLD missing */ ATTR_NOTHROW_LEAF_LIST
3544#undef ATTR_COLD_NORETURN_NOTHROW_LEAF_LIST
3545#define ATTR_COLD_NORETURN_NOTHROW_LEAF_LIST \
3546 /* ECF_COLD missing */ ATTR_NORETURN_NOTHROW_LEAF_LIST
3547#undef ATTR_COLD_CONST_NORETURN_NOTHROW_LEAF_LIST
3548#define ATTR_COLD_CONST_NORETURN_NOTHROW_LEAF_LIST \
3549 /* ECF_COLD missing */ ATTR_CONST_NORETURN_NOTHROW_LEAF_LIST
3550#undef ATTR_PURE_NOTHROW_LEAF_LIST
3551#define ATTR_PURE_NOTHROW_LEAF_LIST ECF_PURE | ATTR_NOTHROW_LEAF_LIST
3552#undef DEF_BUILTIN_STUB
3553#define DEF_BUILTIN_STUB(ENUM, NAME)
3554#undef DEF_SANITIZER_BUILTIN_1
3555#define DEF_SANITIZER_BUILTIN_1(ENUM, NAME, TYPE, ATTRS) \
3556 do { \
3557 decl = add_builtin_function ("__builtin_" NAME, TYPE, ENUM, \
3558 BUILT_IN_NORMAL, NAME, NULL_TREE); \
3559 set_call_expr_flags (decl, ATTRS); \
3560 set_builtin_decl (ENUM, decl, true); \
3561 } while (0)
3562#undef DEF_SANITIZER_BUILTIN
3563#define DEF_SANITIZER_BUILTIN(ENUM, NAME, TYPE, ATTRS) \
3564 DEF_SANITIZER_BUILTIN_1 (ENUM, NAME, TYPE, ATTRS);
3565
3566#include "sanitizer.def"
3567
3568 /* -fsanitize=object-size uses __builtin_dynamic_object_size and
3569 __builtin_object_size, but they might not be available for e.g. Fortran at
3570 this point. We use DEF_SANITIZER_BUILTIN here only as a convenience
3571 macro. */
3572 if (flag_sanitize & SANITIZE_OBJECT_SIZE)
3573 {
3574 if (!builtin_decl_implicit_p (fncode: BUILT_IN_OBJECT_SIZE))
3575 DEF_SANITIZER_BUILTIN_1 (BUILT_IN_OBJECT_SIZE, "object_size",
3576 BT_FN_SIZE_CONST_PTR_INT,
3577 ATTR_PURE_NOTHROW_LEAF_LIST);
3578 if (!builtin_decl_implicit_p (fncode: BUILT_IN_DYNAMIC_OBJECT_SIZE))
3579 DEF_SANITIZER_BUILTIN_1 (BUILT_IN_DYNAMIC_OBJECT_SIZE,
3580 "dynamic_object_size",
3581 BT_FN_SIZE_CONST_PTR_INT,
3582 ATTR_PURE_NOTHROW_LEAF_LIST);
3583 }
3584
3585#undef DEF_SANITIZER_BUILTIN_1
3586#undef DEF_SANITIZER_BUILTIN
3587#undef DEF_BUILTIN_STUB
3588}
3589
3590/* Called via htab_traverse. Count number of emitted
3591 STRING_CSTs in the constant hash table. */
3592
3593int
3594count_string_csts (constant_descriptor_tree **slot,
3595 unsigned HOST_WIDE_INT *data)
3596{
3597 struct constant_descriptor_tree *desc = *slot;
3598 if (TREE_CODE (desc->value) == STRING_CST
3599 && TREE_ASM_WRITTEN (desc->value)
3600 && asan_protect_global (decl: desc->value))
3601 ++*data;
3602 return 1;
3603}
3604
3605/* Helper structure to pass two parameters to
3606 add_string_csts. */
3607
3608struct asan_add_string_csts_data
3609{
3610 tree type;
3611 vec<constructor_elt, va_gc> *v;
3612};
3613
3614/* Called via hash_table::traverse. Call asan_add_global
3615 on emitted STRING_CSTs from the constant hash table. */
3616
3617int
3618add_string_csts (constant_descriptor_tree **slot,
3619 asan_add_string_csts_data *aascd)
3620{
3621 struct constant_descriptor_tree *desc = *slot;
3622 if (TREE_CODE (desc->value) == STRING_CST
3623 && TREE_ASM_WRITTEN (desc->value)
3624 && asan_protect_global (decl: desc->value))
3625 {
3626 asan_add_global (SYMBOL_REF_DECL (XEXP (desc->rtl, 0)),
3627 type: aascd->type, v: aascd->v);
3628 }
3629 return 1;
3630}
3631
3632/* Needs to be GTY(()), because cgraph_build_static_cdtor may
3633 invoke ggc_collect. */
3634static GTY(()) tree asan_ctor_statements;
3635
3636/* Module-level instrumentation.
3637 - Insert __asan_init_vN() into the list of CTORs.
3638 - TODO: insert redzones around globals.
3639 */
3640
3641void
3642asan_finish_file (void)
3643{
3644 varpool_node *vnode;
3645 unsigned HOST_WIDE_INT gcount = 0;
3646
3647 if (shadow_ptr_types[0] == NULL_TREE)
3648 asan_init_shadow_ptr_types ();
3649 /* Avoid instrumenting code in the asan ctors/dtors.
3650 We don't need to insert padding after the description strings,
3651 nor after .LASAN* array. */
3652 flag_sanitize &= ~SANITIZE_ADDRESS;
3653
3654 /* For user-space we want asan constructors to run first.
3655 Linux kernel does not support priorities other than default, and the only
3656 other user of constructors is coverage. So we run with the default
3657 priority. */
3658 int priority = flag_sanitize & SANITIZE_USER_ADDRESS
3659 ? MAX_RESERVED_INIT_PRIORITY - 1 : DEFAULT_INIT_PRIORITY;
3660
3661 if (flag_sanitize & SANITIZE_USER_ADDRESS)
3662 {
3663 tree fn = builtin_decl_implicit (fncode: BUILT_IN_ASAN_INIT);
3664 append_to_statement_list (build_call_expr (fn, 0), &asan_ctor_statements);
3665 fn = builtin_decl_implicit (fncode: BUILT_IN_ASAN_VERSION_MISMATCH_CHECK);
3666 append_to_statement_list (build_call_expr (fn, 0), &asan_ctor_statements);
3667 }
3668 FOR_EACH_DEFINED_VARIABLE (vnode)
3669 if (TREE_ASM_WRITTEN (vnode->decl)
3670 && asan_protect_global (decl: vnode->decl))
3671 ++gcount;
3672 hash_table<tree_descriptor_hasher> *const_desc_htab = constant_pool_htab ();
3673 const_desc_htab->traverse<unsigned HOST_WIDE_INT *, count_string_csts>
3674 (argument: &gcount);
3675 if (gcount)
3676 {
3677 tree type = asan_global_struct (), var, ctor;
3678 tree dtor_statements = NULL_TREE;
3679 vec<constructor_elt, va_gc> *v;
3680 char buf[20];
3681
3682 type = build_array_type_nelts (type, gcount);
3683 ASM_GENERATE_INTERNAL_LABEL (buf, "LASAN", 0);
3684 var = build_decl (UNKNOWN_LOCATION, VAR_DECL, get_identifier (buf),
3685 type);
3686 TREE_STATIC (var) = 1;
3687 TREE_PUBLIC (var) = 0;
3688 DECL_ARTIFICIAL (var) = 1;
3689 DECL_IGNORED_P (var) = 1;
3690 vec_alloc (v, nelems: gcount);
3691 FOR_EACH_DEFINED_VARIABLE (vnode)
3692 if (TREE_ASM_WRITTEN (vnode->decl)
3693 && asan_protect_global (decl: vnode->decl))
3694 asan_add_global (decl: vnode->decl, TREE_TYPE (type), v);
3695 struct asan_add_string_csts_data aascd;
3696 aascd.type = TREE_TYPE (type);
3697 aascd.v = v;
3698 const_desc_htab->traverse<asan_add_string_csts_data *, add_string_csts>
3699 (argument: &aascd);
3700 ctor = build_constructor (type, v);
3701 TREE_CONSTANT (ctor) = 1;
3702 TREE_STATIC (ctor) = 1;
3703 DECL_INITIAL (var) = ctor;
3704 SET_DECL_ALIGN (var, MAX (DECL_ALIGN (var),
3705 ASAN_SHADOW_GRANULARITY * BITS_PER_UNIT));
3706
3707 varpool_node::finalize_decl (decl: var);
3708
3709 tree fn = builtin_decl_implicit (fncode: BUILT_IN_ASAN_REGISTER_GLOBALS);
3710 tree gcount_tree = build_int_cst (pointer_sized_int_node, gcount);
3711 append_to_statement_list (build_call_expr (fn, 2,
3712 build_fold_addr_expr (var),
3713 gcount_tree),
3714 &asan_ctor_statements);
3715
3716 fn = builtin_decl_implicit (fncode: BUILT_IN_ASAN_UNREGISTER_GLOBALS);
3717 append_to_statement_list (build_call_expr (fn, 2,
3718 build_fold_addr_expr (var),
3719 gcount_tree),
3720 &dtor_statements);
3721 cgraph_build_static_cdtor (which: 'D', body: dtor_statements, priority);
3722 }
3723 if (asan_ctor_statements)
3724 cgraph_build_static_cdtor (which: 'I', body: asan_ctor_statements, priority);
3725 flag_sanitize |= SANITIZE_ADDRESS;
3726}
3727
3728/* Poison or unpoison (depending on IS_CLOBBER variable) shadow memory based
3729 on SHADOW address. Newly added statements will be added to ITER with
3730 given location LOC. We mark SIZE bytes in shadow memory, where
3731 LAST_CHUNK_SIZE is greater than zero in situation where we are at the
3732 end of a variable. */
3733
3734static void
3735asan_store_shadow_bytes (gimple_stmt_iterator *iter, location_t loc,
3736 tree shadow,
3737 unsigned HOST_WIDE_INT base_addr_offset,
3738 bool is_clobber, unsigned size,
3739 unsigned last_chunk_size)
3740{
3741 tree shadow_ptr_type;
3742
3743 switch (size)
3744 {
3745 case 1:
3746 shadow_ptr_type = shadow_ptr_types[0];
3747 break;
3748 case 2:
3749 shadow_ptr_type = shadow_ptr_types[1];
3750 break;
3751 case 4:
3752 shadow_ptr_type = shadow_ptr_types[2];
3753 break;
3754 default:
3755 gcc_unreachable ();
3756 }
3757
3758 unsigned char c = (char) is_clobber ? ASAN_STACK_MAGIC_USE_AFTER_SCOPE : 0;
3759 unsigned HOST_WIDE_INT val = 0;
3760 unsigned last_pos = size;
3761 if (last_chunk_size && !is_clobber)
3762 last_pos = BYTES_BIG_ENDIAN ? 0 : size - 1;
3763 for (unsigned i = 0; i < size; ++i)
3764 {
3765 unsigned char shadow_c = c;
3766 if (i == last_pos)
3767 shadow_c = last_chunk_size;
3768 val |= (unsigned HOST_WIDE_INT) shadow_c << (BITS_PER_UNIT * i);
3769 }
3770
3771 /* Handle last chunk in unpoisoning. */
3772 tree magic = build_int_cst (TREE_TYPE (shadow_ptr_type), val);
3773
3774 tree dest = build2 (MEM_REF, TREE_TYPE (shadow_ptr_type), shadow,
3775 build_int_cst (shadow_ptr_type, base_addr_offset));
3776
3777 gimple *g = gimple_build_assign (dest, magic);
3778 gimple_set_location (g, location: loc);
3779 gsi_insert_after (iter, g, GSI_NEW_STMT);
3780}
3781
3782/* Expand the ASAN_MARK builtins. */
3783
3784bool
3785asan_expand_mark_ifn (gimple_stmt_iterator *iter)
3786{
3787 gimple *g = gsi_stmt (i: *iter);
3788 location_t loc = gimple_location (g);
3789 HOST_WIDE_INT flag = tree_to_shwi (gimple_call_arg (gs: g, index: 0));
3790 bool is_poison = ((asan_mark_flags)flag) == ASAN_MARK_POISON;
3791
3792 tree base = gimple_call_arg (gs: g, index: 1);
3793 gcc_checking_assert (TREE_CODE (base) == ADDR_EXPR);
3794 tree decl = TREE_OPERAND (base, 0);
3795
3796 /* For a nested function, we can have: ASAN_MARK (2, &FRAME.2.fp_input, 4) */
3797 if (TREE_CODE (decl) == COMPONENT_REF
3798 && DECL_NONLOCAL_FRAME (TREE_OPERAND (decl, 0)))
3799 decl = TREE_OPERAND (decl, 0);
3800
3801 gcc_checking_assert (TREE_CODE (decl) == VAR_DECL);
3802
3803 if (hwasan_sanitize_p ())
3804 {
3805 gcc_assert (param_hwasan_instrument_stack);
3806 gimple_seq stmts = NULL;
3807 /* Here we swap ASAN_MARK calls for HWASAN_MARK.
3808 This is because we are using the approach of using ASAN_MARK as a
3809 synonym until here.
3810 That approach means we don't yet have to duplicate all the special
3811 cases for ASAN_MARK and ASAN_POISON with the exact same handling but
3812 called HWASAN_MARK etc.
3813
3814 N.b. __asan_poison_stack_memory (which implements ASAN_MARK for ASAN)
3815 rounds the size up to its shadow memory granularity, while
3816 __hwasan_tag_memory (which implements the same for HWASAN) does not.
3817 Hence we emit HWASAN_MARK with an aligned size unlike ASAN_MARK. */
3818 tree len = gimple_call_arg (gs: g, index: 2);
3819 tree new_len = gimple_build_round_up (seq: &stmts, loc, size_type_node, old_size: len,
3820 HWASAN_TAG_GRANULE_SIZE);
3821 gimple_build (seq: &stmts, loc, fn: CFN_HWASAN_MARK,
3822 void_type_node, args: gimple_call_arg (gs: g, index: 0),
3823 args: base, args: new_len);
3824 gsi_replace_with_seq (iter, stmts, true);
3825 return false;
3826 }
3827
3828 if (is_poison)
3829 {
3830 if (asan_handled_variables == NULL)
3831 asan_handled_variables = new hash_set<tree> (16);
3832 asan_handled_variables->add (k: decl);
3833 }
3834 tree len = gimple_call_arg (gs: g, index: 2);
3835
3836 gcc_assert (poly_int_tree_p (len));
3837
3838 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
3839 NOP_EXPR, base);
3840 gimple_set_location (g, location: loc);
3841 gsi_replace (iter, g, false);
3842 tree base_addr = gimple_assign_lhs (gs: g);
3843
3844 /* Generate direct emission if size_in_bytes is small. */
3845 unsigned threshold = param_use_after_scope_direct_emission_threshold;
3846 if (tree_fits_uhwi_p (len) && tree_to_uhwi (len) <= threshold)
3847 {
3848 unsigned HOST_WIDE_INT size_in_bytes = tree_to_uhwi (len);
3849 const unsigned HOST_WIDE_INT shadow_size
3850 = shadow_mem_size (size: size_in_bytes);
3851 const unsigned int shadow_align
3852 = (get_pointer_alignment (base) / BITS_PER_UNIT) >> ASAN_SHADOW_SHIFT;
3853
3854 tree shadow = build_shadow_mem_access (gsi: iter, location: loc, base_addr,
3855 shadow_ptr_type: shadow_ptr_types[0], return_address: true);
3856
3857 for (unsigned HOST_WIDE_INT offset = 0; offset < shadow_size;)
3858 {
3859 unsigned size = 1;
3860 if (shadow_size - offset >= 4
3861 && (!STRICT_ALIGNMENT || shadow_align >= 4))
3862 size = 4;
3863 else if (shadow_size - offset >= 2
3864 && (!STRICT_ALIGNMENT || shadow_align >= 2))
3865 size = 2;
3866
3867 unsigned HOST_WIDE_INT last_chunk_size = 0;
3868 unsigned HOST_WIDE_INT s = (offset + size) * ASAN_SHADOW_GRANULARITY;
3869 if (s > size_in_bytes)
3870 last_chunk_size = ASAN_SHADOW_GRANULARITY - (s - size_in_bytes);
3871
3872 asan_store_shadow_bytes (iter, loc, shadow, base_addr_offset: offset, is_clobber: is_poison,
3873 size, last_chunk_size);
3874 offset += size;
3875 }
3876 }
3877 else
3878 {
3879 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
3880 NOP_EXPR, len);
3881 gimple_set_location (g, location: loc);
3882 gsi_safe_insert_before (iter, g);
3883 tree sz_arg = gimple_assign_lhs (gs: g);
3884
3885 tree fun
3886 = builtin_decl_implicit (fncode: is_poison ? BUILT_IN_ASAN_POISON_STACK_MEMORY
3887 : BUILT_IN_ASAN_UNPOISON_STACK_MEMORY);
3888 g = gimple_build_call (fun, 2, base_addr, sz_arg);
3889 gimple_set_location (g, location: loc);
3890 gsi_insert_after (iter, g, GSI_NEW_STMT);
3891 }
3892
3893 return false;
3894}
3895
3896/* Expand the ASAN_{LOAD,STORE} builtins. */
3897
3898bool
3899asan_expand_check_ifn (gimple_stmt_iterator *iter, bool use_calls)
3900{
3901 gcc_assert (!hwasan_sanitize_p ());
3902 gimple *g = gsi_stmt (i: *iter);
3903 location_t loc = gimple_location (g);
3904 bool recover_p;
3905 if (flag_sanitize & SANITIZE_USER_ADDRESS)
3906 recover_p = (flag_sanitize_recover & SANITIZE_USER_ADDRESS) != 0;
3907 else
3908 recover_p = (flag_sanitize_recover & SANITIZE_KERNEL_ADDRESS) != 0;
3909
3910 HOST_WIDE_INT flags = tree_to_shwi (gimple_call_arg (gs: g, index: 0));
3911 gcc_assert (flags < ASAN_CHECK_LAST);
3912 bool is_scalar_access = (flags & ASAN_CHECK_SCALAR_ACCESS) != 0;
3913 bool is_store = (flags & ASAN_CHECK_STORE) != 0;
3914 bool is_non_zero_len = (flags & ASAN_CHECK_NON_ZERO_LEN) != 0;
3915
3916 tree base = gimple_call_arg (gs: g, index: 1);
3917 tree len = gimple_call_arg (gs: g, index: 2);
3918 HOST_WIDE_INT align = tree_to_shwi (gimple_call_arg (gs: g, index: 3));
3919
3920 HOST_WIDE_INT size_in_bytes
3921 = is_scalar_access && tree_fits_shwi_p (len) ? tree_to_shwi (len) : -1;
3922
3923 if (use_calls)
3924 {
3925 /* Instrument using callbacks. */
3926 gimple *g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
3927 NOP_EXPR, base);
3928 gimple_set_location (g, location: loc);
3929 gsi_insert_before (iter, g, GSI_SAME_STMT);
3930 tree base_addr = gimple_assign_lhs (gs: g);
3931
3932 int nargs;
3933 tree fun = check_func (is_store, recover_p, size_in_bytes, nargs: &nargs);
3934 if (nargs == 1)
3935 g = gimple_build_call (fun, 1, base_addr);
3936 else
3937 {
3938 gcc_assert (nargs == 2);
3939 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
3940 NOP_EXPR, len);
3941 gimple_set_location (g, location: loc);
3942 gsi_insert_before (iter, g, GSI_SAME_STMT);
3943 tree sz_arg = gimple_assign_lhs (gs: g);
3944 g = gimple_build_call (fun, nargs, base_addr, sz_arg);
3945 }
3946 gimple_set_location (g, location: loc);
3947 gsi_replace (iter, g, false);
3948 return false;
3949 }
3950
3951 HOST_WIDE_INT real_size_in_bytes = size_in_bytes == -1 ? 1 : size_in_bytes;
3952
3953 tree shadow_ptr_type = shadow_ptr_types[real_size_in_bytes == 16 ? 1 : 0];
3954 tree shadow_type = TREE_TYPE (shadow_ptr_type);
3955
3956 gimple_stmt_iterator gsi = *iter;
3957
3958 if (!is_non_zero_len)
3959 {
3960 /* So, the length of the memory area to asan-protect is
3961 non-constant. Let's guard the generated instrumentation code
3962 like:
3963
3964 if (len != 0)
3965 {
3966 //asan instrumentation code goes here.
3967 }
3968 // falltrough instructions, starting with *ITER. */
3969
3970 g = gimple_build_cond (NE_EXPR,
3971 len,
3972 build_int_cst (TREE_TYPE (len), 0),
3973 NULL_TREE, NULL_TREE);
3974 gimple_set_location (g, location: loc);
3975
3976 basic_block then_bb, fallthrough_bb;
3977 insert_if_then_before_iter (cond: as_a <gcond *> (p: g), iter,
3978 /*then_more_likely_p=*/true,
3979 then_bb: &then_bb, fallthrough_bb: &fallthrough_bb);
3980 /* Note that fallthrough_bb starts with the statement that was
3981 pointed to by ITER. */
3982
3983 /* The 'then block' of the 'if (len != 0) condition is where
3984 we'll generate the asan instrumentation code now. */
3985 gsi = gsi_last_bb (bb: then_bb);
3986 }
3987
3988 /* Get an iterator on the point where we can add the condition
3989 statement for the instrumentation. */
3990 basic_block then_bb, else_bb;
3991 gsi = create_cond_insert_point (iter: &gsi, /*before_p*/false,
3992 /*then_more_likely_p=*/false,
3993 /*create_then_fallthru_edge*/recover_p,
3994 then_block: &then_bb,
3995 fallthrough_block: &else_bb);
3996
3997 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
3998 NOP_EXPR, base);
3999 gimple_set_location (g, location: loc);
4000 gsi_insert_before (&gsi, g, GSI_NEW_STMT);
4001 tree base_addr = gimple_assign_lhs (gs: g);
4002
4003 tree t = NULL_TREE;
4004 if (real_size_in_bytes >= 8)
4005 {
4006 tree shadow = build_shadow_mem_access (gsi: &gsi, location: loc, base_addr,
4007 shadow_ptr_type);
4008 t = shadow;
4009 }
4010 else
4011 {
4012 /* Slow path for 1, 2 and 4 byte accesses. */
4013 /* Test (shadow != 0)
4014 & ((base_addr & 7) + (real_size_in_bytes - 1)) >= shadow). */
4015 tree shadow = build_shadow_mem_access (gsi: &gsi, location: loc, base_addr,
4016 shadow_ptr_type);
4017 gimple *shadow_test = build_assign (NE_EXPR, shadow, 0);
4018 gimple_seq seq = NULL;
4019 gimple_seq_add_stmt (&seq, shadow_test);
4020 /* Aligned (>= 8 bytes) can test just
4021 (real_size_in_bytes - 1 >= shadow), as base_addr & 7 is known
4022 to be 0. */
4023 if (align < 8)
4024 {
4025 gimple_seq_add_stmt (&seq, build_assign (BIT_AND_EXPR,
4026 base_addr, 7));
4027 gimple_seq_add_stmt (&seq,
4028 build_type_cast (shadow_type,
4029 gimple_seq_last (s: seq)));
4030 if (real_size_in_bytes > 1)
4031 gimple_seq_add_stmt (&seq,
4032 build_assign (PLUS_EXPR,
4033 gimple_seq_last (s: seq),
4034 real_size_in_bytes - 1));
4035 t = gimple_assign_lhs (gs: gimple_seq_last_stmt (s: seq));
4036 }
4037 else
4038 t = build_int_cst (shadow_type, real_size_in_bytes - 1);
4039 gimple_seq_add_stmt (&seq, build_assign (GE_EXPR, t, shadow));
4040 gimple_seq_add_stmt (&seq, build_assign (BIT_AND_EXPR, shadow_test,
4041 gimple_seq_last (s: seq)));
4042 t = gimple_assign_lhs (gs: gimple_seq_last (s: seq));
4043 gimple_seq_set_location (seq, loc);
4044 gsi_insert_seq_after (&gsi, seq, GSI_CONTINUE_LINKING);
4045
4046 /* For non-constant, misaligned or otherwise weird access sizes,
4047 check first and last byte. */
4048 if (size_in_bytes == -1)
4049 {
4050 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
4051 MINUS_EXPR, len,
4052 build_int_cst (pointer_sized_int_node, 1));
4053 gimple_set_location (g, location: loc);
4054 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
4055 tree last = gimple_assign_lhs (gs: g);
4056 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
4057 PLUS_EXPR, base_addr, last);
4058 gimple_set_location (g, location: loc);
4059 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
4060 tree base_end_addr = gimple_assign_lhs (gs: g);
4061
4062 tree shadow = build_shadow_mem_access (gsi: &gsi, location: loc, base_addr: base_end_addr,
4063 shadow_ptr_type);
4064 gimple *shadow_test = build_assign (NE_EXPR, shadow, 0);
4065 gimple_seq seq = NULL;
4066 gimple_seq_add_stmt (&seq, shadow_test);
4067 gimple_seq_add_stmt (&seq, build_assign (BIT_AND_EXPR,
4068 base_end_addr, 7));
4069 gimple_seq_add_stmt (&seq, build_type_cast (shadow_type,
4070 gimple_seq_last (s: seq)));
4071 gimple_seq_add_stmt (&seq, build_assign (GE_EXPR,
4072 gimple_seq_last (s: seq),
4073 shadow));
4074 gimple_seq_add_stmt (&seq, build_assign (BIT_AND_EXPR, shadow_test,
4075 gimple_seq_last (s: seq)));
4076 gimple_seq_add_stmt (&seq, build_assign (BIT_IOR_EXPR, t,
4077 gimple_seq_last (s: seq)));
4078 t = gimple_assign_lhs (gs: gimple_seq_last (s: seq));
4079 gimple_seq_set_location (seq, loc);
4080 gsi_insert_seq_after (&gsi, seq, GSI_CONTINUE_LINKING);
4081 }
4082 }
4083
4084 g = gimple_build_cond (NE_EXPR, t, build_int_cst (TREE_TYPE (t), 0),
4085 NULL_TREE, NULL_TREE);
4086 gimple_set_location (g, location: loc);
4087 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
4088
4089 /* Generate call to the run-time library (e.g. __asan_report_load8). */
4090 gsi = gsi_start_bb (bb: then_bb);
4091 int nargs;
4092 tree fun = report_error_func (is_store, recover_p, size_in_bytes, nargs: &nargs);
4093 g = gimple_build_call (fun, nargs, base_addr, len);
4094 gimple_set_location (g, location: loc);
4095 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
4096
4097 gsi_remove (iter, true);
4098 *iter = gsi_start_bb (bb: else_bb);
4099
4100 return true;
4101}
4102
4103/* Create ASAN shadow variable for a VAR_DECL which has been rewritten
4104 into SSA. Already seen VAR_DECLs are stored in SHADOW_VARS_MAPPING. */
4105
4106static tree
4107create_asan_shadow_var (tree var_decl,
4108 hash_map<tree, tree> &shadow_vars_mapping)
4109{
4110 tree *slot = shadow_vars_mapping.get (k: var_decl);
4111 if (slot == NULL)
4112 {
4113 tree shadow_var = copy_node (var_decl);
4114
4115 copy_body_data id;
4116 memset (s: &id, c: 0, n: sizeof (copy_body_data));
4117 id.src_fn = id.dst_fn = current_function_decl;
4118 copy_decl_for_dup_finish (id: &id, decl: var_decl, copy: shadow_var);
4119
4120 DECL_ARTIFICIAL (shadow_var) = 1;
4121 DECL_IGNORED_P (shadow_var) = 1;
4122 DECL_SEEN_IN_BIND_EXPR_P (shadow_var) = 0;
4123 gimple_add_tmp_var (shadow_var);
4124
4125 shadow_vars_mapping.put (k: var_decl, v: shadow_var);
4126 return shadow_var;
4127 }
4128 else
4129 return *slot;
4130}
4131
4132/* Expand ASAN_POISON ifn. */
4133
4134bool
4135asan_expand_poison_ifn (gimple_stmt_iterator *iter,
4136 bool *need_commit_edge_insert,
4137 hash_map<tree, tree> &shadow_vars_mapping)
4138{
4139 gimple *g = gsi_stmt (i: *iter);
4140 tree poisoned_var = gimple_call_lhs (gs: g);
4141 if (!poisoned_var || has_zero_uses (var: poisoned_var))
4142 {
4143 gsi_remove (iter, true);
4144 return true;
4145 }
4146
4147 if (SSA_NAME_VAR (poisoned_var) == NULL_TREE)
4148 SET_SSA_NAME_VAR_OR_IDENTIFIER (poisoned_var,
4149 create_tmp_var (TREE_TYPE (poisoned_var)));
4150
4151 tree shadow_var = create_asan_shadow_var (SSA_NAME_VAR (poisoned_var),
4152 shadow_vars_mapping);
4153
4154 bool recover_p;
4155 if (flag_sanitize & SANITIZE_USER_ADDRESS)
4156 recover_p = (flag_sanitize_recover & SANITIZE_USER_ADDRESS) != 0;
4157 else
4158 recover_p = (flag_sanitize_recover & SANITIZE_KERNEL_ADDRESS) != 0;
4159 tree size = DECL_SIZE_UNIT (shadow_var);
4160 gimple *poison_call
4161 = gimple_build_call_internal (IFN_ASAN_MARK, 3,
4162 build_int_cst (integer_type_node,
4163 ASAN_MARK_POISON),
4164 build_fold_addr_expr (shadow_var), size);
4165
4166 gimple *use;
4167 imm_use_iterator imm_iter;
4168 FOR_EACH_IMM_USE_STMT (use, imm_iter, poisoned_var)
4169 {
4170 if (is_gimple_debug (gs: use))
4171 continue;
4172
4173 int nargs;
4174 bool store_p = gimple_call_internal_p (gs: use, fn: IFN_ASAN_POISON_USE);
4175 gcall *call;
4176 if (hwasan_sanitize_p ())
4177 {
4178 tree fun = builtin_decl_implicit (fncode: BUILT_IN_HWASAN_TAG_MISMATCH4);
4179 /* NOTE: hwasan has no __hwasan_report_* functions like asan does.
4180 We use __hwasan_tag_mismatch4 with arguments that tell it the
4181 size of access and load to report all tag mismatches.
4182
4183 The arguments to this function are:
4184 Address of invalid access.
4185 Bitfield containing information about the access
4186 (access_info)
4187 Pointer to a frame of registers
4188 (for use in printing the contents of registers in a dump)
4189 Not used yet -- to be used by inline instrumentation.
4190 Size of access.
4191
4192 The access_info bitfield encodes the following pieces of
4193 information:
4194 - Is this a store or load?
4195 access_info & 0x10 => store
4196 - Should the program continue after reporting the error?
4197 access_info & 0x20 => recover
4198 - What size access is this (not used here since we can always
4199 pass the size in the last argument)
4200
4201 if (access_info & 0xf == 0xf)
4202 size is taken from last argument.
4203 else
4204 size == 1 << (access_info & 0xf)
4205
4206 The last argument contains the size of the access iff the
4207 access_info size indicator is 0xf (we always use this argument
4208 rather than storing the size in the access_info bitfield).
4209
4210 See the function definition `__hwasan_tag_mismatch4` in
4211 libsanitizer/hwasan for the full definition.
4212 */
4213 unsigned access_info = (0x20 * recover_p)
4214 + (0x10 * store_p)
4215 + (0xf);
4216 call = gimple_build_call (fun, 4,
4217 build_fold_addr_expr (shadow_var),
4218 build_int_cst (pointer_sized_int_node,
4219 access_info),
4220 build_int_cst (pointer_sized_int_node, 0),
4221 size);
4222 }
4223 else
4224 {
4225 tree fun = report_error_func (is_store: store_p, recover_p, size_in_bytes: tree_to_uhwi (size),
4226 nargs: &nargs);
4227 call = gimple_build_call (fun, 1,
4228 build_fold_addr_expr (shadow_var));
4229 }
4230 gimple_set_location (g: call, location: gimple_location (g: use));
4231 gimple *call_to_insert = call;
4232
4233 /* The USE can be a gimple PHI node. If so, insert the call on
4234 all edges leading to the PHI node. */
4235 if (is_a <gphi *> (p: use))
4236 {
4237 gphi *phi = dyn_cast<gphi *> (p: use);
4238 for (unsigned i = 0; i < gimple_phi_num_args (gs: phi); ++i)
4239 if (gimple_phi_arg_def (gs: phi, index: i) == poisoned_var)
4240 {
4241 edge e = gimple_phi_arg_edge (phi, i);
4242
4243 /* Do not insert on an edge we can't split. */
4244 if (e->flags & EDGE_ABNORMAL)
4245 continue;
4246
4247 if (call_to_insert == NULL)
4248 call_to_insert = gimple_copy (call);
4249
4250 gsi_insert_seq_on_edge (e, call_to_insert);
4251 *need_commit_edge_insert = true;
4252 call_to_insert = NULL;
4253 }
4254 }
4255 else
4256 {
4257 gimple_stmt_iterator gsi = gsi_for_stmt (use);
4258 if (store_p)
4259 gsi_replace (&gsi, call, true);
4260 else
4261 gsi_insert_before (&gsi, call, GSI_NEW_STMT);
4262 }
4263 }
4264
4265 SSA_NAME_IS_DEFAULT_DEF (poisoned_var) = true;
4266 SSA_NAME_DEF_STMT (poisoned_var) = gimple_build_nop ();
4267 gsi_replace (iter, poison_call, false);
4268
4269 return true;
4270}
4271
4272/* Instrument the current function. */
4273
4274static unsigned int
4275asan_instrument (void)
4276{
4277 if (hwasan_sanitize_p ())
4278 {
4279 transform_statements ();
4280 return 0;
4281 }
4282
4283 if (shadow_ptr_types[0] == NULL_TREE)
4284 asan_init_shadow_ptr_types ();
4285 transform_statements ();
4286 last_alloca_addr = NULL_TREE;
4287 return 0;
4288}
4289
4290static bool
4291gate_asan (void)
4292{
4293 return sanitize_flags_p (flag: SANITIZE_ADDRESS);
4294}
4295
4296namespace {
4297
4298const pass_data pass_data_asan =
4299{
4300 .type: GIMPLE_PASS, /* type */
4301 .name: "asan", /* name */
4302 .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */
4303 .tv_id: TV_NONE, /* tv_id */
4304 .properties_required: ( PROP_ssa | PROP_cfg | PROP_gimple_leh ), /* properties_required */
4305 .properties_provided: 0, /* properties_provided */
4306 .properties_destroyed: 0, /* properties_destroyed */
4307 .todo_flags_start: 0, /* todo_flags_start */
4308 TODO_update_ssa, /* todo_flags_finish */
4309};
4310
4311class pass_asan : public gimple_opt_pass
4312{
4313public:
4314 pass_asan (gcc::context *ctxt)
4315 : gimple_opt_pass (pass_data_asan, ctxt)
4316 {}
4317
4318 /* opt_pass methods: */
4319 opt_pass * clone () final override { return new pass_asan (m_ctxt); }
4320 bool gate (function *) final override
4321 {
4322 return gate_asan () || gate_hwasan ();
4323 }
4324 unsigned int execute (function *) final override
4325 {
4326 return asan_instrument ();
4327 }
4328
4329}; // class pass_asan
4330
4331} // anon namespace
4332
4333gimple_opt_pass *
4334make_pass_asan (gcc::context *ctxt)
4335{
4336 return new pass_asan (ctxt);
4337}
4338
4339namespace {
4340
4341const pass_data pass_data_asan_O0 =
4342{
4343 .type: GIMPLE_PASS, /* type */
4344 .name: "asan0", /* name */
4345 .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */
4346 .tv_id: TV_NONE, /* tv_id */
4347 .properties_required: ( PROP_ssa | PROP_cfg | PROP_gimple_leh ), /* properties_required */
4348 .properties_provided: 0, /* properties_provided */
4349 .properties_destroyed: 0, /* properties_destroyed */
4350 .todo_flags_start: 0, /* todo_flags_start */
4351 TODO_update_ssa, /* todo_flags_finish */
4352};
4353
4354class pass_asan_O0 : public gimple_opt_pass
4355{
4356public:
4357 pass_asan_O0 (gcc::context *ctxt)
4358 : gimple_opt_pass (pass_data_asan_O0, ctxt)
4359 {}
4360
4361 /* opt_pass methods: */
4362 bool gate (function *) final override
4363 {
4364 return !optimize && (gate_asan () || gate_hwasan ());
4365 }
4366 unsigned int execute (function *) final override
4367 {
4368 return asan_instrument ();
4369 }
4370
4371}; // class pass_asan_O0
4372
4373} // anon namespace
4374
4375gimple_opt_pass *
4376make_pass_asan_O0 (gcc::context *ctxt)
4377{
4378 return new pass_asan_O0 (ctxt);
4379}
4380
4381/* HWASAN */
4382
4383/* For stack tagging:
4384
4385 Return the offset from the frame base tag that the "next" expanded object
4386 should have. */
4387uint8_t
4388hwasan_current_frame_tag ()
4389{
4390 return hwasan_frame_tag_offset;
4391}
4392
4393/* For stack tagging:
4394
4395 Return the 'base pointer' for this function. If that base pointer has not
4396 yet been created then we create a register to hold it and record the insns
4397 to initialize the register in `hwasan_frame_base_init_seq` for later
4398 emission. */
4399rtx
4400hwasan_frame_base ()
4401{
4402 if (! hwasan_frame_base_ptr)
4403 {
4404 start_sequence ();
4405 hwasan_frame_base_ptr
4406 = force_reg (Pmode,
4407 targetm.memtag.insert_random_tag (virtual_stack_vars_rtx,
4408 NULL_RTX));
4409 hwasan_frame_base_init_seq = get_insns ();
4410 end_sequence ();
4411 }
4412
4413 return hwasan_frame_base_ptr;
4414}
4415
4416/* For stack tagging:
4417
4418 Check whether this RTX is a standard pointer addressing the base of the
4419 stack variables for this frame. Returns true if the RTX is either
4420 virtual_stack_vars_rtx or hwasan_frame_base_ptr. */
4421bool
4422stack_vars_base_reg_p (rtx base)
4423{
4424 return base == virtual_stack_vars_rtx || base == hwasan_frame_base_ptr;
4425}
4426
4427/* For stack tagging:
4428
4429 Emit frame base initialisation.
4430 If hwasan_frame_base has been used before here then
4431 hwasan_frame_base_init_seq contains the sequence of instructions to
4432 initialize it. This must be put just before the hwasan prologue, so we emit
4433 the insns before parm_birth_insn (which will point to the first instruction
4434 of the hwasan prologue if it exists).
4435
4436 We update `parm_birth_insn` to point to the start of this initialisation
4437 since that represents the end of the initialisation done by
4438 expand_function_{start,end} functions and we want to maintain that. */
4439void
4440hwasan_maybe_emit_frame_base_init ()
4441{
4442 if (! hwasan_frame_base_init_seq)
4443 return;
4444 emit_insn_before (hwasan_frame_base_init_seq, parm_birth_insn);
4445 parm_birth_insn = hwasan_frame_base_init_seq;
4446}
4447
4448/* Record a compile-time constant size stack variable that HWASAN will need to
4449 tag. This record of the range of a stack variable will be used by
4450 `hwasan_emit_prologue` to emit the RTL at the start of each frame which will
4451 set tags in the shadow memory according to the assigned tag for each object.
4452
4453 The range that the object spans in stack space should be described by the
4454 bounds `untagged_base + nearest_offset` and
4455 `untagged_base + farthest_offset`.
4456 `tagged_base` is the base address which contains the "base frame tag" for
4457 this frame, and from which the value to address this object with will be
4458 calculated.
4459
4460 We record the `untagged_base` since the functions in the hwasan library we
4461 use to tag memory take pointers without a tag. */
4462void
4463hwasan_record_stack_var (rtx untagged_base, rtx tagged_base,
4464 poly_int64 nearest_offset, poly_int64 farthest_offset)
4465{
4466 hwasan_stack_var cur_var;
4467 cur_var.untagged_base = untagged_base;
4468 cur_var.tagged_base = tagged_base;
4469 cur_var.nearest_offset = nearest_offset;
4470 cur_var.farthest_offset = farthest_offset;
4471 cur_var.tag_offset = hwasan_current_frame_tag ();
4472
4473 hwasan_tagged_stack_vars.safe_push (obj: cur_var);
4474}
4475
4476/* Return the RTX representing the farthest extent of the statically allocated
4477 stack objects for this frame. If hwasan_frame_base_ptr has not been
4478 initialized then we are not storing any static variables on the stack in
4479 this frame. In this case we return NULL_RTX to represent that.
4480
4481 Otherwise simply return virtual_stack_vars_rtx + frame_offset. */
4482rtx
4483hwasan_get_frame_extent ()
4484{
4485 return (hwasan_frame_base_ptr
4486 ? plus_constant (Pmode, virtual_stack_vars_rtx, frame_offset)
4487 : NULL_RTX);
4488}
4489
4490/* For stack tagging:
4491
4492 Increment the frame tag offset modulo the size a tag can represent. */
4493void
4494hwasan_increment_frame_tag ()
4495{
4496 uint8_t tag_bits = HWASAN_TAG_SIZE;
4497 gcc_assert (HWASAN_TAG_SIZE
4498 <= sizeof (hwasan_frame_tag_offset) * CHAR_BIT);
4499 hwasan_frame_tag_offset = (hwasan_frame_tag_offset + 1) % (1 << tag_bits);
4500 /* The "background tag" of the stack is zero by definition.
4501 This is the tag that objects like parameters passed on the stack and
4502 spilled registers are given. It is handy to avoid this tag for objects
4503 whose tags we decide ourselves, partly to ensure that buffer overruns
4504 can't affect these important variables (e.g. saved link register, saved
4505 stack pointer etc) and partly to make debugging easier (everything with a
4506 tag of zero is space allocated automatically by the compiler).
4507
4508 This is not feasible when using random frame tags (the default
4509 configuration for hwasan) since the tag for the given frame is randomly
4510 chosen at runtime. In order to avoid any tags matching the stack
4511 background we would need to decide tag offsets at runtime instead of
4512 compile time (and pay the resulting performance cost).
4513
4514 When not using random base tags for each frame (i.e. when compiled with
4515 `--param hwasan-random-frame-tag=0`) the base tag for each frame is zero.
4516 This means the tag that each object gets is equal to the
4517 hwasan_frame_tag_offset used in determining it.
4518 When this is the case we *can* ensure no object gets the tag of zero by
4519 simply ensuring no object has the hwasan_frame_tag_offset of zero.
4520
4521 There is the extra complication that we only record the
4522 hwasan_frame_tag_offset here (which is the offset from the tag stored in
4523 the stack pointer). In the kernel, the tag in the stack pointer is 0xff
4524 rather than zero. This does not cause problems since tags of 0xff are
4525 never checked in the kernel. As mentioned at the beginning of this
4526 comment the background tag of the stack is zero by definition, which means
4527 that for the kernel we should skip offsets of both 0 and 1 from the stack
4528 pointer. Avoiding the offset of 0 ensures we use a tag which will be
4529 checked, avoiding the offset of 1 ensures we use a tag that is not the
4530 same as the background. */
4531 if (hwasan_frame_tag_offset == 0 && ! param_hwasan_random_frame_tag)
4532 hwasan_frame_tag_offset += 1;
4533 if (hwasan_frame_tag_offset == 1 && ! param_hwasan_random_frame_tag
4534 && sanitize_flags_p (flag: SANITIZE_KERNEL_HWADDRESS))
4535 hwasan_frame_tag_offset += 1;
4536}
4537
4538/* Clear internal state for the next function.
4539 This function is called before variables on the stack get expanded, in
4540 `init_vars_expansion`. */
4541void
4542hwasan_record_frame_init ()
4543{
4544 delete asan_used_labels;
4545 asan_used_labels = NULL;
4546
4547 /* If this isn't the case then some stack variable was recorded *before*
4548 hwasan_record_frame_init is called, yet *after* the hwasan prologue for
4549 the previous frame was emitted. Such stack variables would not have
4550 their shadow stack filled in. */
4551 gcc_assert (hwasan_tagged_stack_vars.is_empty ());
4552 hwasan_frame_base_ptr = NULL_RTX;
4553 hwasan_frame_base_init_seq = NULL;
4554
4555 /* When not using a random frame tag we can avoid the background stack
4556 color which gives the user a little better debug output upon a crash.
4557 Meanwhile, when using a random frame tag it will be nice to avoid adding
4558 tags for the first object since that is unnecessary extra work.
4559 Hence set the initial hwasan_frame_tag_offset to be 0 if using a random
4560 frame tag and 1 otherwise.
4561
4562 As described in hwasan_increment_frame_tag, in the kernel the stack
4563 pointer has the tag 0xff. That means that to avoid 0xff and 0 (the tag
4564 which the kernel does not check and the background tag respectively) we
4565 start with a tag offset of 2. */
4566 hwasan_frame_tag_offset = param_hwasan_random_frame_tag
4567 ? 0
4568 : sanitize_flags_p (flag: SANITIZE_KERNEL_HWADDRESS) ? 2 : 1;
4569}
4570
4571/* For stack tagging:
4572 (Emits HWASAN equivalent of what is emitted by
4573 `asan_emit_stack_protection`).
4574
4575 Emits the extra prologue code to set the shadow stack as required for HWASAN
4576 stack instrumentation.
4577
4578 Uses the vector of recorded stack variables hwasan_tagged_stack_vars. When
4579 this function has completed hwasan_tagged_stack_vars is empty and all
4580 objects it had pointed to are deallocated. */
4581void
4582hwasan_emit_prologue ()
4583{
4584 /* We need untagged base pointers since libhwasan only accepts untagged
4585 pointers in __hwasan_tag_memory. We need the tagged base pointer to obtain
4586 the base tag for an offset. */
4587
4588 if (hwasan_tagged_stack_vars.is_empty ())
4589 return;
4590
4591 poly_int64 bot = 0, top = 0;
4592 for (hwasan_stack_var &cur : hwasan_tagged_stack_vars)
4593 {
4594 poly_int64 nearest = cur.nearest_offset;
4595 poly_int64 farthest = cur.farthest_offset;
4596
4597 if (known_ge (nearest, farthest))
4598 {
4599 top = nearest;
4600 bot = farthest;
4601 }
4602 else
4603 {
4604 /* Given how these values are calculated, one must be known greater
4605 than the other. */
4606 gcc_assert (known_le (nearest, farthest));
4607 top = farthest;
4608 bot = nearest;
4609 }
4610 poly_int64 size = (top - bot);
4611
4612 /* Assert the edge of each variable is aligned to the HWASAN tag granule
4613 size. */
4614 gcc_assert (multiple_p (top, HWASAN_TAG_GRANULE_SIZE));
4615 gcc_assert (multiple_p (bot, HWASAN_TAG_GRANULE_SIZE));
4616 gcc_assert (multiple_p (size, HWASAN_TAG_GRANULE_SIZE));
4617
4618 rtx fn = init_one_libfunc ("__hwasan_tag_memory");
4619 rtx base_tag = targetm.memtag.extract_tag (cur.tagged_base, NULL_RTX);
4620 rtx tag = plus_constant (QImode, base_tag, cur.tag_offset);
4621 tag = hwasan_truncate_to_tag_size (tag, NULL_RTX);
4622
4623 rtx bottom = convert_memory_address (ptr_mode,
4624 plus_constant (Pmode,
4625 cur.untagged_base,
4626 bot));
4627 emit_library_call (fun: fn, fn_type: LCT_NORMAL, VOIDmode,
4628 arg1: bottom, arg1_mode: ptr_mode,
4629 arg2: tag, QImode,
4630 arg3: gen_int_mode (size, ptr_mode), arg3_mode: ptr_mode);
4631 }
4632 /* Clear the stack vars, we've emitted the prologue for them all now. */
4633 hwasan_tagged_stack_vars.truncate (size: 0);
4634}
4635
4636/* For stack tagging:
4637
4638 Return RTL insns to clear the tags between DYNAMIC and VARS pointers
4639 into the stack. These instructions should be emitted at the end of
4640 every function.
4641
4642 If `dynamic` is NULL_RTX then no insns are returned. */
4643rtx_insn *
4644hwasan_emit_untag_frame (rtx dynamic, rtx vars)
4645{
4646 if (! dynamic)
4647 return NULL;
4648
4649 start_sequence ();
4650
4651 dynamic = convert_memory_address (ptr_mode, dynamic);
4652 vars = convert_memory_address (ptr_mode, vars);
4653
4654 rtx top_rtx;
4655 rtx bot_rtx;
4656 if (FRAME_GROWS_DOWNWARD)
4657 {
4658 top_rtx = vars;
4659 bot_rtx = dynamic;
4660 }
4661 else
4662 {
4663 top_rtx = dynamic;
4664 bot_rtx = vars;
4665 }
4666
4667 rtx size_rtx = expand_simple_binop (ptr_mode, MINUS, top_rtx, bot_rtx,
4668 NULL_RTX, /* unsignedp = */0,
4669 OPTAB_DIRECT);
4670
4671 rtx fn = init_one_libfunc ("__hwasan_tag_memory");
4672 emit_library_call (fun: fn, fn_type: LCT_NORMAL, VOIDmode,
4673 arg1: bot_rtx, arg1_mode: ptr_mode,
4674 HWASAN_STACK_BACKGROUND, QImode,
4675 arg3: size_rtx, arg3_mode: ptr_mode);
4676
4677 do_pending_stack_adjust ();
4678 rtx_insn *insns = get_insns ();
4679 end_sequence ();
4680 return insns;
4681}
4682
4683/* Needs to be GTY(()), because cgraph_build_static_cdtor may
4684 invoke ggc_collect. */
4685static GTY(()) tree hwasan_ctor_statements;
4686
4687/* Insert module initialization into this TU. This initialization calls the
4688 initialization code for libhwasan. */
4689void
4690hwasan_finish_file (void)
4691{
4692 /* Do not emit constructor initialization for the kernel.
4693 (the kernel has its own initialization already). */
4694 if (flag_sanitize & SANITIZE_KERNEL_HWADDRESS)
4695 return;
4696
4697 /* Avoid instrumenting code in the hwasan constructors/destructors. */
4698 flag_sanitize &= ~SANITIZE_HWADDRESS;
4699 int priority = MAX_RESERVED_INIT_PRIORITY - 1;
4700 tree fn = builtin_decl_implicit (fncode: BUILT_IN_HWASAN_INIT);
4701 append_to_statement_list (build_call_expr (fn, 0), &hwasan_ctor_statements);
4702 cgraph_build_static_cdtor (which: 'I', body: hwasan_ctor_statements, priority);
4703 flag_sanitize |= SANITIZE_HWADDRESS;
4704}
4705
4706/* For stack tagging:
4707
4708 Truncate `tag` to the number of bits that a tag uses (i.e. to
4709 HWASAN_TAG_SIZE). Store the result in `target` if it's convenient. */
4710rtx
4711hwasan_truncate_to_tag_size (rtx tag, rtx target)
4712{
4713 gcc_assert (GET_MODE (tag) == QImode);
4714 if (HWASAN_TAG_SIZE != GET_MODE_PRECISION (QImode))
4715 {
4716 gcc_assert (GET_MODE_PRECISION (QImode) > HWASAN_TAG_SIZE);
4717 rtx mask = gen_int_mode ((HOST_WIDE_INT_1U << HWASAN_TAG_SIZE) - 1,
4718 QImode);
4719 tag = expand_simple_binop (QImode, AND, tag, mask, target,
4720 /* unsignedp = */1, OPTAB_WIDEN);
4721 gcc_assert (tag);
4722 }
4723 return tag;
4724}
4725
4726/* Construct a function tree for __hwasan_{load,store}{1,2,4,8,16,_n}.
4727 IS_STORE is either 1 (for a store) or 0 (for a load). */
4728static combined_fn
4729hwasan_check_func (bool is_store, bool recover_p, HOST_WIDE_INT size_in_bytes,
4730 int *nargs)
4731{
4732 static enum built_in_function check[2][2][6]
4733 = { { { BUILT_IN_HWASAN_LOAD1, BUILT_IN_HWASAN_LOAD2,
4734 BUILT_IN_HWASAN_LOAD4, BUILT_IN_HWASAN_LOAD8,
4735 BUILT_IN_HWASAN_LOAD16, BUILT_IN_HWASAN_LOADN },
4736 { BUILT_IN_HWASAN_STORE1, BUILT_IN_HWASAN_STORE2,
4737 BUILT_IN_HWASAN_STORE4, BUILT_IN_HWASAN_STORE8,
4738 BUILT_IN_HWASAN_STORE16, BUILT_IN_HWASAN_STOREN } },
4739 { { BUILT_IN_HWASAN_LOAD1_NOABORT,
4740 BUILT_IN_HWASAN_LOAD2_NOABORT,
4741 BUILT_IN_HWASAN_LOAD4_NOABORT,
4742 BUILT_IN_HWASAN_LOAD8_NOABORT,
4743 BUILT_IN_HWASAN_LOAD16_NOABORT,
4744 BUILT_IN_HWASAN_LOADN_NOABORT },
4745 { BUILT_IN_HWASAN_STORE1_NOABORT,
4746 BUILT_IN_HWASAN_STORE2_NOABORT,
4747 BUILT_IN_HWASAN_STORE4_NOABORT,
4748 BUILT_IN_HWASAN_STORE8_NOABORT,
4749 BUILT_IN_HWASAN_STORE16_NOABORT,
4750 BUILT_IN_HWASAN_STOREN_NOABORT } } };
4751 if (size_in_bytes == -1)
4752 {
4753 *nargs = 2;
4754 return as_combined_fn (fn: check[recover_p][is_store][5]);
4755 }
4756 *nargs = 1;
4757 int size_log2 = exact_log2 (x: size_in_bytes);
4758 gcc_assert (size_log2 >= 0 && size_log2 <= 5);
4759 return as_combined_fn (fn: check[recover_p][is_store][size_log2]);
4760}
4761
4762/* Expand the HWASAN_{LOAD,STORE} builtins. */
4763bool
4764hwasan_expand_check_ifn (gimple_stmt_iterator *iter, bool)
4765{
4766 gimple *g = gsi_stmt (i: *iter);
4767 location_t loc = gimple_location (g);
4768 bool recover_p;
4769 if (flag_sanitize & SANITIZE_USER_HWADDRESS)
4770 recover_p = (flag_sanitize_recover & SANITIZE_USER_HWADDRESS) != 0;
4771 else
4772 recover_p = (flag_sanitize_recover & SANITIZE_KERNEL_HWADDRESS) != 0;
4773
4774 HOST_WIDE_INT flags = tree_to_shwi (gimple_call_arg (gs: g, index: 0));
4775 gcc_assert (flags < ASAN_CHECK_LAST);
4776 bool is_scalar_access = (flags & ASAN_CHECK_SCALAR_ACCESS) != 0;
4777 bool is_store = (flags & ASAN_CHECK_STORE) != 0;
4778 bool is_non_zero_len = (flags & ASAN_CHECK_NON_ZERO_LEN) != 0;
4779
4780 tree base = gimple_call_arg (gs: g, index: 1);
4781 tree len = gimple_call_arg (gs: g, index: 2);
4782
4783 /* `align` is unused for HWASAN_CHECK, but we pass the argument anyway
4784 since that way the arguments match ASAN_CHECK. */
4785 /* HOST_WIDE_INT align = tree_to_shwi (gimple_call_arg (g, 3)); */
4786
4787 unsigned HOST_WIDE_INT size_in_bytes
4788 = is_scalar_access ? tree_to_shwi (len) : -1;
4789
4790 gimple_stmt_iterator gsi = *iter;
4791
4792 if (!is_non_zero_len)
4793 {
4794 /* So, the length of the memory area to hwasan-protect is
4795 non-constant. Let's guard the generated instrumentation code
4796 like:
4797
4798 if (len != 0)
4799 {
4800 // hwasan instrumentation code goes here.
4801 }
4802 // falltrough instructions, starting with *ITER. */
4803
4804 g = gimple_build_cond (NE_EXPR,
4805 len,
4806 build_int_cst (TREE_TYPE (len), 0),
4807 NULL_TREE, NULL_TREE);
4808 gimple_set_location (g, location: loc);
4809
4810 basic_block then_bb, fallthrough_bb;
4811 insert_if_then_before_iter (cond: as_a <gcond *> (p: g), iter,
4812 /*then_more_likely_p=*/true,
4813 then_bb: &then_bb, fallthrough_bb: &fallthrough_bb);
4814 /* Note that fallthrough_bb starts with the statement that was
4815 pointed to by ITER. */
4816
4817 /* The 'then block' of the 'if (len != 0) condition is where
4818 we'll generate the hwasan instrumentation code now. */
4819 gsi = gsi_last_bb (bb: then_bb);
4820 }
4821
4822 gimple_seq stmts = NULL;
4823 tree base_addr = gimple_build (seq: &stmts, loc, code: NOP_EXPR,
4824 pointer_sized_int_node, ops: base);
4825
4826 int nargs = 0;
4827 combined_fn fn
4828 = hwasan_check_func (is_store, recover_p, size_in_bytes, nargs: &nargs);
4829 if (nargs == 1)
4830 gimple_build (seq: &stmts, loc, fn, void_type_node, args: base_addr);
4831 else
4832 {
4833 gcc_assert (nargs == 2);
4834 tree sz_arg = gimple_build (seq: &stmts, loc, code: NOP_EXPR,
4835 pointer_sized_int_node, ops: len);
4836 gimple_build (seq: &stmts, loc, fn, void_type_node, args: base_addr, args: sz_arg);
4837 }
4838
4839 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
4840 gsi_remove (iter, true);
4841 *iter = gsi;
4842 return false;
4843}
4844
4845/* For stack tagging:
4846
4847 Dummy: the HWASAN_MARK internal function should only ever be in the code
4848 after the sanopt pass. */
4849bool
4850hwasan_expand_mark_ifn (gimple_stmt_iterator *)
4851{
4852 gcc_unreachable ();
4853}
4854
4855bool
4856gate_hwasan ()
4857{
4858 return hwasan_sanitize_p ();
4859}
4860
4861#include "gt-asan.h"
4862

source code of gcc/asan.cc