1/* AddressSanitizer, a fast memory error detector.
2 Copyright (C) 2012-2023 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) && gimple_store_p (gs: stmt))
1376 {
1377 asan_mem_ref r;
1378 asan_mem_ref_init (ref: &r, NULL, access_size: 1);
1379
1380 r.start = gimple_call_lhs (gs: stmt);
1381 r.access_size = int_size_in_bytes (TREE_TYPE (r.start));
1382 return has_mem_ref_been_instrumented (ref: &r);
1383 }
1384
1385 return false;
1386}
1387
1388/* Insert a memory reference into the hash table. */
1389
1390static void
1391update_mem_ref_hash_table (tree ref, HOST_WIDE_INT access_size)
1392{
1393 hash_table<asan_mem_ref_hasher> *ht = get_mem_ref_hash_table ();
1394
1395 asan_mem_ref r;
1396 asan_mem_ref_init (ref: &r, start: ref, access_size);
1397
1398 asan_mem_ref **slot = ht->find_slot (value: &r, insert: INSERT);
1399 if (*slot == NULL || (*slot)->access_size < access_size)
1400 *slot = asan_mem_ref_new (start: ref, access_size);
1401}
1402
1403/* Initialize shadow_ptr_types array. */
1404
1405static void
1406asan_init_shadow_ptr_types (void)
1407{
1408 asan_shadow_set = new_alias_set ();
1409 tree types[3] = { signed_char_type_node, short_integer_type_node,
1410 integer_type_node };
1411
1412 for (unsigned i = 0; i < 3; i++)
1413 {
1414 shadow_ptr_types[i] = build_distinct_type_copy (types[i]);
1415 TYPE_ALIAS_SET (shadow_ptr_types[i]) = asan_shadow_set;
1416 shadow_ptr_types[i] = build_pointer_type (shadow_ptr_types[i]);
1417 }
1418
1419 initialize_sanitizer_builtins ();
1420}
1421
1422/* Create ADDR_EXPR of STRING_CST with the PP pretty printer text. */
1423
1424static tree
1425asan_pp_string (pretty_printer *pp)
1426{
1427 const char *buf = pp_formatted_text (pp);
1428 size_t len = strlen (s: buf);
1429 tree ret = build_string (len + 1, buf);
1430 TREE_TYPE (ret)
1431 = build_array_type (TREE_TYPE (shadow_ptr_types[0]),
1432 build_index_type (size_int (len)));
1433 TREE_READONLY (ret) = 1;
1434 TREE_STATIC (ret) = 1;
1435 return build1 (ADDR_EXPR, shadow_ptr_types[0], ret);
1436}
1437
1438/* Clear shadow memory at SHADOW_MEM, LEN bytes. Can't call a library call here
1439 though. */
1440
1441static void
1442asan_clear_shadow (rtx shadow_mem, HOST_WIDE_INT len)
1443{
1444 rtx_insn *insn, *insns, *jump;
1445 rtx_code_label *top_label;
1446 rtx end, addr, tmp;
1447
1448 gcc_assert ((len & 3) == 0);
1449 start_sequence ();
1450 clear_storage (shadow_mem, GEN_INT (len), BLOCK_OP_NORMAL);
1451 insns = get_insns ();
1452 end_sequence ();
1453 for (insn = insns; insn; insn = NEXT_INSN (insn))
1454 if (CALL_P (insn))
1455 break;
1456 if (insn == NULL_RTX)
1457 {
1458 emit_insn (insns);
1459 return;
1460 }
1461
1462 top_label = gen_label_rtx ();
1463 addr = copy_to_mode_reg (Pmode, XEXP (shadow_mem, 0));
1464 shadow_mem = adjust_automodify_address (shadow_mem, SImode, addr, 0);
1465 end = force_reg (Pmode, plus_constant (Pmode, addr, len));
1466 emit_label (top_label);
1467
1468 emit_move_insn (shadow_mem, const0_rtx);
1469 tmp = expand_simple_binop (Pmode, PLUS, addr, gen_int_mode (4, Pmode), addr,
1470 true, OPTAB_LIB_WIDEN);
1471 if (tmp != addr)
1472 emit_move_insn (addr, tmp);
1473 emit_cmp_and_jump_insns (addr, end, LT, NULL_RTX, Pmode, true, top_label);
1474 jump = get_last_insn ();
1475 gcc_assert (JUMP_P (jump));
1476 add_reg_br_prob_note (jump,
1477 profile_probability::guessed_always ()
1478 .apply_scale (num: 80, den: 100));
1479}
1480
1481void
1482asan_function_start (void)
1483{
1484 section *fnsec = function_section (current_function_decl);
1485 switch_to_section (fnsec);
1486 ASM_OUTPUT_DEBUG_LABEL (asm_out_file, "LASANPC",
1487 current_function_funcdef_no);
1488}
1489
1490/* Return number of shadow bytes that are occupied by a local variable
1491 of SIZE bytes. */
1492
1493static unsigned HOST_WIDE_INT
1494shadow_mem_size (unsigned HOST_WIDE_INT size)
1495{
1496 /* It must be possible to align stack variables to granularity
1497 of shadow memory. */
1498 gcc_assert (BITS_PER_UNIT
1499 * ASAN_SHADOW_GRANULARITY <= MAX_SUPPORTED_STACK_ALIGNMENT);
1500
1501 return ROUND_UP (size, ASAN_SHADOW_GRANULARITY) / ASAN_SHADOW_GRANULARITY;
1502}
1503
1504/* Always emit 4 bytes at a time. */
1505#define RZ_BUFFER_SIZE 4
1506
1507/* ASAN redzone buffer container that handles emission of shadow bytes. */
1508class asan_redzone_buffer
1509{
1510public:
1511 /* Constructor. */
1512 asan_redzone_buffer (rtx shadow_mem, HOST_WIDE_INT prev_offset):
1513 m_shadow_mem (shadow_mem), m_prev_offset (prev_offset),
1514 m_original_offset (prev_offset), m_shadow_bytes (RZ_BUFFER_SIZE)
1515 {}
1516
1517 /* Emit VALUE shadow byte at a given OFFSET. */
1518 void emit_redzone_byte (HOST_WIDE_INT offset, unsigned char value);
1519
1520 /* Emit RTX emission of the content of the buffer. */
1521 void flush_redzone_payload (void);
1522
1523private:
1524 /* Flush if the content of the buffer is full
1525 (equal to RZ_BUFFER_SIZE). */
1526 void flush_if_full (void);
1527
1528 /* Memory where we last emitted a redzone payload. */
1529 rtx m_shadow_mem;
1530
1531 /* Relative offset where we last emitted a redzone payload. */
1532 HOST_WIDE_INT m_prev_offset;
1533
1534 /* Relative original offset. Used for checking only. */
1535 HOST_WIDE_INT m_original_offset;
1536
1537public:
1538 /* Buffer with redzone payload. */
1539 auto_vec<unsigned char> m_shadow_bytes;
1540};
1541
1542/* Emit VALUE shadow byte at a given OFFSET. */
1543
1544void
1545asan_redzone_buffer::emit_redzone_byte (HOST_WIDE_INT offset,
1546 unsigned char value)
1547{
1548 gcc_assert ((offset & (ASAN_SHADOW_GRANULARITY - 1)) == 0);
1549 gcc_assert (offset >= m_prev_offset);
1550
1551 HOST_WIDE_INT off
1552 = m_prev_offset + ASAN_SHADOW_GRANULARITY * m_shadow_bytes.length ();
1553 if (off == offset)
1554 /* Consecutive shadow memory byte. */;
1555 else if (offset < m_prev_offset + (HOST_WIDE_INT) (ASAN_SHADOW_GRANULARITY
1556 * RZ_BUFFER_SIZE)
1557 && !m_shadow_bytes.is_empty ())
1558 {
1559 /* Shadow memory byte with a small gap. */
1560 for (; off < offset; off += ASAN_SHADOW_GRANULARITY)
1561 m_shadow_bytes.safe_push (obj: 0);
1562 }
1563 else
1564 {
1565 if (!m_shadow_bytes.is_empty ())
1566 flush_redzone_payload ();
1567
1568 /* Maybe start earlier in order to use aligned store. */
1569 HOST_WIDE_INT align = (offset - m_prev_offset) % ASAN_RED_ZONE_SIZE;
1570 if (align)
1571 {
1572 offset -= align;
1573 for (unsigned i = 0; i < align / BITS_PER_UNIT; i++)
1574 m_shadow_bytes.safe_push (obj: 0);
1575 }
1576
1577 /* Adjust m_prev_offset and m_shadow_mem. */
1578 HOST_WIDE_INT diff = offset - m_prev_offset;
1579 m_shadow_mem = adjust_address (m_shadow_mem, VOIDmode,
1580 diff >> ASAN_SHADOW_SHIFT);
1581 m_prev_offset = offset;
1582 }
1583 m_shadow_bytes.safe_push (obj: value);
1584 flush_if_full ();
1585}
1586
1587/* Emit RTX emission of the content of the buffer. */
1588
1589void
1590asan_redzone_buffer::flush_redzone_payload (void)
1591{
1592 gcc_assert (WORDS_BIG_ENDIAN == BYTES_BIG_ENDIAN);
1593
1594 if (m_shadow_bytes.is_empty ())
1595 return;
1596
1597 /* Be sure we always emit to an aligned address. */
1598 gcc_assert (((m_prev_offset - m_original_offset)
1599 & (ASAN_RED_ZONE_SIZE - 1)) == 0);
1600
1601 /* Fill it to RZ_BUFFER_SIZE bytes with zeros if needed. */
1602 unsigned l = m_shadow_bytes.length ();
1603 for (unsigned i = 0; i <= RZ_BUFFER_SIZE - l; i++)
1604 m_shadow_bytes.safe_push (obj: 0);
1605
1606 if (dump_file && (dump_flags & TDF_DETAILS))
1607 fprintf (stream: dump_file,
1608 format: "Flushing rzbuffer at offset %" PRId64 " with: ", m_prev_offset);
1609
1610 unsigned HOST_WIDE_INT val = 0;
1611 for (unsigned i = 0; i < RZ_BUFFER_SIZE; i++)
1612 {
1613 unsigned char v
1614 = m_shadow_bytes[BYTES_BIG_ENDIAN ? RZ_BUFFER_SIZE - i - 1 : i];
1615 val |= (unsigned HOST_WIDE_INT)v << (BITS_PER_UNIT * i);
1616 if (dump_file && (dump_flags & TDF_DETAILS))
1617 fprintf (stream: dump_file, format: "%02x ", v);
1618 }
1619
1620 if (dump_file && (dump_flags & TDF_DETAILS))
1621 fprintf (stream: dump_file, format: "\n");
1622
1623 rtx c = gen_int_mode (val, SImode);
1624 m_shadow_mem = adjust_address (m_shadow_mem, SImode, 0);
1625 emit_move_insn (m_shadow_mem, c);
1626 m_shadow_bytes.truncate (size: 0);
1627}
1628
1629/* Flush if the content of the buffer is full
1630 (equal to RZ_BUFFER_SIZE). */
1631
1632void
1633asan_redzone_buffer::flush_if_full (void)
1634{
1635 if (m_shadow_bytes.length () == RZ_BUFFER_SIZE)
1636 flush_redzone_payload ();
1637}
1638
1639
1640/* HWAddressSanitizer (hwasan) is a probabilistic method for detecting
1641 out-of-bounds and use-after-free bugs.
1642 Read more:
1643 http://code.google.com/p/address-sanitizer/
1644
1645 Similar to AddressSanitizer (asan) it consists of two parts: the
1646 instrumentation module in this file, and a run-time library.
1647
1648 The instrumentation module adds a run-time check before every memory insn in
1649 the same manner as asan (see the block comment for AddressSanitizer above).
1650 Currently, hwasan only adds out-of-line instrumentation, where each check is
1651 implemented as a function call to the run-time library. Hence a check for a
1652 load of N bytes from address X would be implemented with a function call to
1653 __hwasan_loadN(X), and checking a store of N bytes from address X would be
1654 implemented with a function call to __hwasan_storeN(X).
1655
1656 The main difference between hwasan and asan is in the information stored to
1657 help this checking. Both sanitizers use a shadow memory area which stores
1658 data recording the state of main memory at a corresponding address.
1659
1660 For hwasan, each 16 byte granule in main memory has a corresponding 1 byte
1661 in shadow memory. This shadow address can be calculated with equation:
1662 (addr >> log_2(HWASAN_TAG_GRANULE_SIZE))
1663 + __hwasan_shadow_memory_dynamic_address;
1664 The conversion between real and shadow memory for asan is given in the block
1665 comment at the top of this file.
1666 The description of how this shadow memory is laid out for asan is in the
1667 block comment at the top of this file, here we describe how this shadow
1668 memory is used for hwasan.
1669
1670 For hwasan, each variable is assigned a byte-sized 'tag'. The extent of
1671 the shadow memory for that variable is filled with the assigned tag, and
1672 every pointer referencing that variable has its top byte set to the same
1673 tag. The run-time library redefines malloc so that every allocation returns
1674 a tagged pointer and tags the corresponding shadow memory with the same tag.
1675
1676 On each pointer dereference the tag found in the pointer is compared to the
1677 tag found in the shadow memory corresponding to the accessed memory address.
1678 If these tags are found to differ then this memory access is judged to be
1679 invalid and a report is generated.
1680
1681 This method of bug detection is not perfect -- it can not catch every bad
1682 access -- but catches them probabilistically instead. There is always the
1683 possibility that an invalid memory access will happen to access memory
1684 tagged with the same tag as the pointer that this access used.
1685 The chances of this are approx. 0.4% for any two uncorrelated objects.
1686
1687 Random tag generation can mitigate this problem by decreasing the
1688 probability that an invalid access will be missed in the same manner over
1689 multiple runs. i.e. if two objects are tagged the same in one run of the
1690 binary they are unlikely to be tagged the same in the next run.
1691 Both heap and stack allocated objects have random tags by default.
1692
1693 [16 byte granule implications]
1694 Since the shadow memory only has a resolution on real memory of 16 bytes,
1695 invalid accesses that are within the same 16 byte granule as a valid
1696 address will not be caught.
1697
1698 There is a "short-granule" feature in the runtime library which does catch
1699 such accesses, but this feature is not implemented for stack objects (since
1700 stack objects are allocated and tagged by compiler instrumentation, and
1701 this feature has not yet been implemented in GCC instrumentation).
1702
1703 Another outcome of this 16 byte resolution is that each tagged object must
1704 be 16 byte aligned. If two objects were to share any 16 byte granule in
1705 memory, then they both would have to be given the same tag, and invalid
1706 accesses to one using a pointer to the other would be undetectable.
1707
1708 [Compiler instrumentation]
1709 Compiler instrumentation ensures that two adjacent buffers on the stack are
1710 given different tags, this means an access to one buffer using a pointer
1711 generated from the other (e.g. through buffer overrun) will have mismatched
1712 tags and be caught by hwasan.
1713
1714 We don't randomly tag every object on the stack, since that would require
1715 keeping many registers to record each tag. Instead we randomly generate a
1716 tag for each function frame, and each new stack object uses a tag offset
1717 from that frame tag.
1718 i.e. each object is tagged as RFT + offset, where RFT is the "random frame
1719 tag" generated for this frame.
1720 This means that randomisation does not peturb the difference between tags
1721 on tagged stack objects within a frame, but this is mitigated by the fact
1722 that objects with the same tag within a frame are very far apart
1723 (approx. 2^HWASAN_TAG_SIZE objects apart).
1724
1725 As a demonstration, using the same example program as in the asan block
1726 comment above:
1727
1728 int
1729 foo ()
1730 {
1731 char a[24] = {0};
1732 int b[2] = {0};
1733
1734 a[5] = 1;
1735 b[1] = 2;
1736
1737 return a[5] + b[1];
1738 }
1739
1740 On AArch64 the stack will be ordered as follows for the above function:
1741
1742 Slot 1/ [24 bytes for variable 'a']
1743 Slot 2/ [8 bytes padding for alignment]
1744 Slot 3/ [8 bytes for variable 'b']
1745 Slot 4/ [8 bytes padding for alignment]
1746
1747 (The padding is there to ensure 16 byte alignment as described in the 16
1748 byte granule implications).
1749
1750 While the shadow memory will be ordered as follows:
1751
1752 - 2 bytes (representing 32 bytes in real memory) tagged with RFT + 1.
1753 - 1 byte (representing 16 bytes in real memory) tagged with RFT + 2.
1754
1755 And any pointer to "a" will have the tag RFT + 1, and any pointer to "b"
1756 will have the tag RFT + 2.
1757
1758 [Top Byte Ignore requirements]
1759 Hwasan requires the ability to store an 8 bit tag in every pointer. There
1760 is no instrumentation done to remove this tag from pointers before
1761 dereferencing, which means the hardware must ignore this tag during memory
1762 accesses.
1763
1764 Architectures where this feature is available should indicate this using
1765 the TARGET_MEMTAG_CAN_TAG_ADDRESSES hook.
1766
1767 [Stack requires cleanup on unwinding]
1768 During normal operation of a hwasan sanitized program more space in the
1769 shadow memory becomes tagged as the stack grows. As the stack shrinks this
1770 shadow memory space must become untagged. If it is not untagged then when
1771 the stack grows again (during other function calls later on in the program)
1772 objects on the stack that are usually not tagged (e.g. parameters passed on
1773 the stack) can be placed in memory whose shadow space is tagged with
1774 something else, and accesses can cause false positive reports.
1775
1776 Hence we place untagging code on every epilogue of functions which tag some
1777 stack objects.
1778
1779 Moreover, the run-time library intercepts longjmp & setjmp to untag when
1780 the stack is unwound this way.
1781
1782 C++ exceptions are not yet handled, which means this sanitizer can not
1783 handle C++ code that throws exceptions -- it will give false positives
1784 after an exception has been thrown. The implementation that the hwasan
1785 library has for handling these relies on the frame pointer being after any
1786 local variables. This is not generally the case for GCC. */
1787
1788
1789/* Returns whether we are tagging pointers and checking those tags on memory
1790 access. */
1791bool
1792hwasan_sanitize_p ()
1793{
1794 return sanitize_flags_p (flag: SANITIZE_HWADDRESS);
1795}
1796
1797/* Are we tagging the stack? */
1798bool
1799hwasan_sanitize_stack_p ()
1800{
1801 return (hwasan_sanitize_p () && param_hwasan_instrument_stack);
1802}
1803
1804/* Are we tagging alloca objects? */
1805bool
1806hwasan_sanitize_allocas_p (void)
1807{
1808 return (hwasan_sanitize_stack_p () && param_hwasan_instrument_allocas);
1809}
1810
1811/* Should we instrument reads? */
1812bool
1813hwasan_instrument_reads (void)
1814{
1815 return (hwasan_sanitize_p () && param_hwasan_instrument_reads);
1816}
1817
1818/* Should we instrument writes? */
1819bool
1820hwasan_instrument_writes (void)
1821{
1822 return (hwasan_sanitize_p () && param_hwasan_instrument_writes);
1823}
1824
1825/* Should we instrument builtin calls? */
1826bool
1827hwasan_memintrin (void)
1828{
1829 return (hwasan_sanitize_p () && param_hwasan_instrument_mem_intrinsics);
1830}
1831
1832/* Insert code to protect stack vars. The prologue sequence should be emitted
1833 directly, epilogue sequence returned. BASE is the register holding the
1834 stack base, against which OFFSETS array offsets are relative to, OFFSETS
1835 array contains pairs of offsets in reverse order, always the end offset
1836 of some gap that needs protection followed by starting offset,
1837 and DECLS is an array of representative decls for each var partition.
1838 LENGTH is the length of the OFFSETS array, DECLS array is LENGTH / 2 - 1
1839 elements long (OFFSETS include gap before the first variable as well
1840 as gaps after each stack variable). PBASE is, if non-NULL, some pseudo
1841 register which stack vars DECL_RTLs are based on. Either BASE should be
1842 assigned to PBASE, when not doing use after return protection, or
1843 corresponding address based on __asan_stack_malloc* return value. */
1844
1845rtx_insn *
1846asan_emit_stack_protection (rtx base, rtx pbase, unsigned int alignb,
1847 HOST_WIDE_INT *offsets, tree *decls, int length)
1848{
1849 rtx shadow_base, shadow_mem, ret, mem, orig_base;
1850 rtx_code_label *lab;
1851 rtx_insn *insns;
1852 char buf[32];
1853 HOST_WIDE_INT base_offset = offsets[length - 1];
1854 HOST_WIDE_INT base_align_bias = 0, offset, prev_offset;
1855 HOST_WIDE_INT asan_frame_size = offsets[0] - base_offset;
1856 HOST_WIDE_INT last_offset, last_size, last_size_aligned;
1857 int l;
1858 unsigned char cur_shadow_byte = ASAN_STACK_MAGIC_LEFT;
1859 tree str_cst, decl, id;
1860 int use_after_return_class = -1;
1861
1862 /* Don't emit anything when doing error recovery, the assertions
1863 might fail e.g. if a function had a frame offset overflow. */
1864 if (seen_error ())
1865 return NULL;
1866
1867 if (shadow_ptr_types[0] == NULL_TREE)
1868 asan_init_shadow_ptr_types ();
1869
1870 expanded_location cfun_xloc
1871 = expand_location (DECL_SOURCE_LOCATION (current_function_decl));
1872
1873 /* First of all, prepare the description string. */
1874 pretty_printer asan_pp;
1875
1876 pp_decimal_int (&asan_pp, length / 2 - 1);
1877 pp_space (&asan_pp);
1878 for (l = length - 2; l; l -= 2)
1879 {
1880 tree decl = decls[l / 2 - 1];
1881 pp_wide_integer (pp: &asan_pp, i: offsets[l] - base_offset);
1882 pp_space (&asan_pp);
1883 pp_wide_integer (pp: &asan_pp, i: offsets[l - 1] - offsets[l]);
1884 pp_space (&asan_pp);
1885
1886 expanded_location xloc
1887 = expand_location (DECL_SOURCE_LOCATION (decl));
1888 char location[32];
1889
1890 if (xloc.file == cfun_xloc.file)
1891 sprintf (s: location, format: ":%d", xloc.line);
1892 else
1893 location[0] = '\0';
1894
1895 if (DECL_P (decl) && DECL_NAME (decl))
1896 {
1897 unsigned idlen
1898 = IDENTIFIER_LENGTH (DECL_NAME (decl)) + strlen (s: location);
1899 pp_decimal_int (&asan_pp, idlen);
1900 pp_space (&asan_pp);
1901 pp_tree_identifier (&asan_pp, DECL_NAME (decl));
1902 pp_string (&asan_pp, location);
1903 }
1904 else
1905 pp_string (&asan_pp, "9 <unknown>");
1906
1907 if (l > 2)
1908 pp_space (&asan_pp);
1909 }
1910 str_cst = asan_pp_string (pp: &asan_pp);
1911
1912 /* Emit the prologue sequence. */
1913 if (asan_frame_size > 32 && asan_frame_size <= 65536 && pbase
1914 && param_asan_use_after_return)
1915 {
1916 use_after_return_class = floor_log2 (x: asan_frame_size - 1) - 5;
1917 /* __asan_stack_malloc_N guarantees alignment
1918 N < 6 ? (64 << N) : 4096 bytes. */
1919 if (alignb > (use_after_return_class < 6
1920 ? (64U << use_after_return_class) : 4096U))
1921 use_after_return_class = -1;
1922 else if (alignb > ASAN_RED_ZONE_SIZE && (asan_frame_size & (alignb - 1)))
1923 base_align_bias = ((asan_frame_size + alignb - 1)
1924 & ~(alignb - HOST_WIDE_INT_1)) - asan_frame_size;
1925 }
1926
1927 /* Align base if target is STRICT_ALIGNMENT. */
1928 if (STRICT_ALIGNMENT)
1929 {
1930 const HOST_WIDE_INT align
1931 = (GET_MODE_ALIGNMENT (SImode) / BITS_PER_UNIT) << ASAN_SHADOW_SHIFT;
1932 base = expand_binop (Pmode, and_optab, base, gen_int_mode (-align, Pmode),
1933 NULL_RTX, 1, OPTAB_DIRECT);
1934 }
1935
1936 if (use_after_return_class == -1 && pbase)
1937 emit_move_insn (pbase, base);
1938
1939 base = expand_binop (Pmode, add_optab, base,
1940 gen_int_mode (base_offset - base_align_bias, Pmode),
1941 NULL_RTX, 1, OPTAB_DIRECT);
1942 orig_base = NULL_RTX;
1943 if (use_after_return_class != -1)
1944 {
1945 if (asan_detect_stack_use_after_return == NULL_TREE)
1946 {
1947 id = get_identifier ("__asan_option_detect_stack_use_after_return");
1948 decl = build_decl (BUILTINS_LOCATION, VAR_DECL, id,
1949 integer_type_node);
1950 SET_DECL_ASSEMBLER_NAME (decl, id);
1951 TREE_ADDRESSABLE (decl) = 1;
1952 DECL_ARTIFICIAL (decl) = 1;
1953 DECL_IGNORED_P (decl) = 1;
1954 DECL_EXTERNAL (decl) = 1;
1955 TREE_STATIC (decl) = 1;
1956 TREE_PUBLIC (decl) = 1;
1957 TREE_USED (decl) = 1;
1958 asan_detect_stack_use_after_return = decl;
1959 }
1960 orig_base = gen_reg_rtx (Pmode);
1961 emit_move_insn (orig_base, base);
1962 ret = expand_normal (exp: asan_detect_stack_use_after_return);
1963 lab = gen_label_rtx ();
1964 emit_cmp_and_jump_insns (ret, const0_rtx, EQ, NULL_RTX,
1965 VOIDmode, 0, lab,
1966 prob: profile_probability::very_likely ());
1967 snprintf (s: buf, maxlen: sizeof buf, format: "__asan_stack_malloc_%d",
1968 use_after_return_class);
1969 ret = init_one_libfunc (buf);
1970 ret = emit_library_call_value (fun: ret, NULL_RTX, fn_type: LCT_NORMAL, outmode: ptr_mode,
1971 GEN_INT (asan_frame_size
1972 + base_align_bias),
1973 TYPE_MODE (pointer_sized_int_node));
1974 /* __asan_stack_malloc_[n] returns a pointer to fake stack if succeeded
1975 and NULL otherwise. Check RET value is NULL here and jump over the
1976 BASE reassignment in this case. Otherwise, reassign BASE to RET. */
1977 emit_cmp_and_jump_insns (ret, const0_rtx, EQ, NULL_RTX,
1978 VOIDmode, 0, lab,
1979 prob: profile_probability:: very_unlikely ());
1980 ret = convert_memory_address (Pmode, ret);
1981 emit_move_insn (base, ret);
1982 emit_label (lab);
1983 emit_move_insn (pbase, expand_binop (Pmode, add_optab, base,
1984 gen_int_mode (base_align_bias
1985 - base_offset, Pmode),
1986 NULL_RTX, 1, OPTAB_DIRECT));
1987 }
1988 mem = gen_rtx_MEM (ptr_mode, base);
1989 mem = adjust_address (mem, VOIDmode, base_align_bias);
1990 emit_move_insn (mem, gen_int_mode (ASAN_STACK_FRAME_MAGIC, ptr_mode));
1991 mem = adjust_address (mem, VOIDmode, GET_MODE_SIZE (ptr_mode));
1992 emit_move_insn (mem, expand_normal (exp: str_cst));
1993 mem = adjust_address (mem, VOIDmode, GET_MODE_SIZE (ptr_mode));
1994 ASM_GENERATE_INTERNAL_LABEL (buf, "LASANPC", current_function_funcdef_no);
1995 id = get_identifier (buf);
1996 decl = build_decl (DECL_SOURCE_LOCATION (current_function_decl),
1997 VAR_DECL, id, char_type_node);
1998 SET_DECL_ASSEMBLER_NAME (decl, id);
1999 TREE_ADDRESSABLE (decl) = 1;
2000 TREE_READONLY (decl) = 1;
2001 DECL_ARTIFICIAL (decl) = 1;
2002 DECL_IGNORED_P (decl) = 1;
2003 TREE_STATIC (decl) = 1;
2004 TREE_PUBLIC (decl) = 0;
2005 TREE_USED (decl) = 1;
2006 DECL_INITIAL (decl) = decl;
2007 TREE_ASM_WRITTEN (decl) = 1;
2008 TREE_ASM_WRITTEN (id) = 1;
2009 emit_move_insn (mem, expand_normal (build_fold_addr_expr (decl)));
2010 shadow_base = expand_binop (Pmode, lshr_optab, base,
2011 gen_int_shift_amount (Pmode, ASAN_SHADOW_SHIFT),
2012 NULL_RTX, 1, OPTAB_DIRECT);
2013 shadow_base
2014 = plus_constant (Pmode, shadow_base,
2015 asan_shadow_offset ()
2016 + (base_align_bias >> ASAN_SHADOW_SHIFT));
2017 gcc_assert (asan_shadow_set != -1
2018 && (ASAN_RED_ZONE_SIZE >> ASAN_SHADOW_SHIFT) == 4);
2019 shadow_mem = gen_rtx_MEM (SImode, shadow_base);
2020 set_mem_alias_set (shadow_mem, asan_shadow_set);
2021 if (STRICT_ALIGNMENT)
2022 set_mem_align (shadow_mem, (GET_MODE_ALIGNMENT (SImode)));
2023 prev_offset = base_offset;
2024
2025 asan_redzone_buffer rz_buffer (shadow_mem, prev_offset);
2026 for (l = length; l; l -= 2)
2027 {
2028 if (l == 2)
2029 cur_shadow_byte = ASAN_STACK_MAGIC_RIGHT;
2030 offset = offsets[l - 1];
2031
2032 bool extra_byte = (offset - base_offset) & (ASAN_SHADOW_GRANULARITY - 1);
2033 /* If a red-zone is not aligned to ASAN_SHADOW_GRANULARITY then
2034 the previous stack variable has size % ASAN_SHADOW_GRANULARITY != 0.
2035 In that case we have to emit one extra byte that will describe
2036 how many bytes (our of ASAN_SHADOW_GRANULARITY) can be accessed. */
2037 if (extra_byte)
2038 {
2039 HOST_WIDE_INT aoff
2040 = base_offset + ((offset - base_offset)
2041 & ~(ASAN_SHADOW_GRANULARITY - HOST_WIDE_INT_1));
2042 rz_buffer.emit_redzone_byte (offset: aoff, value: offset - aoff);
2043 offset = aoff + ASAN_SHADOW_GRANULARITY;
2044 }
2045
2046 /* Calculate size of red zone payload. */
2047 while (offset < offsets[l - 2])
2048 {
2049 rz_buffer.emit_redzone_byte (offset, value: cur_shadow_byte);
2050 offset += ASAN_SHADOW_GRANULARITY;
2051 }
2052
2053 cur_shadow_byte = ASAN_STACK_MAGIC_MIDDLE;
2054 }
2055
2056 /* As the automatic variables are aligned to
2057 ASAN_RED_ZONE_SIZE / ASAN_SHADOW_GRANULARITY, the buffer should be
2058 flushed here. */
2059 gcc_assert (rz_buffer.m_shadow_bytes.is_empty ());
2060
2061 do_pending_stack_adjust ();
2062
2063 /* Construct epilogue sequence. */
2064 start_sequence ();
2065
2066 lab = NULL;
2067 if (use_after_return_class != -1)
2068 {
2069 rtx_code_label *lab2 = gen_label_rtx ();
2070 char c = (char) ASAN_STACK_MAGIC_USE_AFTER_RET;
2071 emit_cmp_and_jump_insns (orig_base, base, EQ, NULL_RTX,
2072 VOIDmode, 0, lab2,
2073 prob: profile_probability::very_likely ());
2074 shadow_mem = gen_rtx_MEM (BLKmode, shadow_base);
2075 set_mem_alias_set (shadow_mem, asan_shadow_set);
2076 mem = gen_rtx_MEM (ptr_mode, base);
2077 mem = adjust_address (mem, VOIDmode, base_align_bias);
2078 emit_move_insn (mem, gen_int_mode (ASAN_STACK_RETIRED_MAGIC, ptr_mode));
2079 unsigned HOST_WIDE_INT sz = asan_frame_size >> ASAN_SHADOW_SHIFT;
2080 if (use_after_return_class < 5
2081 && can_store_by_pieces (sz, builtin_memset_read_str, &c,
2082 BITS_PER_UNIT, true))
2083 {
2084 /* Emit:
2085 memset(ShadowBase, kAsanStackAfterReturnMagic, ShadowSize);
2086 **SavedFlagPtr(FakeStack, class_id) = 0
2087 */
2088 store_by_pieces (shadow_mem, sz, builtin_memset_read_str, &c,
2089 BITS_PER_UNIT, true, RETURN_BEGIN);
2090
2091 unsigned HOST_WIDE_INT offset
2092 = (1 << (use_after_return_class + 6));
2093 offset -= GET_MODE_SIZE (mode: ptr_mode);
2094 mem = gen_rtx_MEM (ptr_mode, base);
2095 mem = adjust_address (mem, ptr_mode, offset);
2096 rtx addr = gen_reg_rtx (ptr_mode);
2097 emit_move_insn (addr, mem);
2098 addr = convert_memory_address (Pmode, addr);
2099 mem = gen_rtx_MEM (QImode, addr);
2100 emit_move_insn (mem, const0_rtx);
2101 }
2102 else if (use_after_return_class >= 5
2103 || !set_storage_via_setmem (shadow_mem,
2104 GEN_INT (sz),
2105 gen_int_mode (c, QImode),
2106 BITS_PER_UNIT, BITS_PER_UNIT,
2107 -1, sz, sz, sz))
2108 {
2109 snprintf (s: buf, maxlen: sizeof buf, format: "__asan_stack_free_%d",
2110 use_after_return_class);
2111 ret = init_one_libfunc (buf);
2112 rtx addr = convert_memory_address (ptr_mode, base);
2113 rtx orig_addr = convert_memory_address (ptr_mode, orig_base);
2114 emit_library_call (fun: ret, fn_type: LCT_NORMAL, outmode: ptr_mode, arg1: addr, arg1_mode: ptr_mode,
2115 GEN_INT (asan_frame_size + base_align_bias),
2116 TYPE_MODE (pointer_sized_int_node),
2117 arg3: orig_addr, arg3_mode: ptr_mode);
2118 }
2119 lab = gen_label_rtx ();
2120 emit_jump (lab);
2121 emit_label (lab2);
2122 }
2123
2124 shadow_mem = gen_rtx_MEM (BLKmode, shadow_base);
2125 set_mem_alias_set (shadow_mem, asan_shadow_set);
2126
2127 if (STRICT_ALIGNMENT)
2128 set_mem_align (shadow_mem, (GET_MODE_ALIGNMENT (SImode)));
2129
2130 prev_offset = base_offset;
2131 last_offset = base_offset;
2132 last_size = 0;
2133 last_size_aligned = 0;
2134 for (l = length; l; l -= 2)
2135 {
2136 offset = base_offset + ((offsets[l - 1] - base_offset)
2137 & ~(ASAN_RED_ZONE_SIZE - HOST_WIDE_INT_1));
2138 if (last_offset + last_size_aligned < offset)
2139 {
2140 shadow_mem = adjust_address (shadow_mem, VOIDmode,
2141 (last_offset - prev_offset)
2142 >> ASAN_SHADOW_SHIFT);
2143 prev_offset = last_offset;
2144 asan_clear_shadow (shadow_mem, len: last_size_aligned >> ASAN_SHADOW_SHIFT);
2145 last_offset = offset;
2146 last_size = 0;
2147 }
2148 else
2149 last_size = offset - last_offset;
2150 last_size += base_offset + ((offsets[l - 2] - base_offset)
2151 & ~(ASAN_MIN_RED_ZONE_SIZE - HOST_WIDE_INT_1))
2152 - offset;
2153
2154 /* Unpoison shadow memory that corresponds to a variable that is
2155 is subject of use-after-return sanitization. */
2156 if (l > 2)
2157 {
2158 decl = decls[l / 2 - 2];
2159 if (asan_handled_variables != NULL
2160 && asan_handled_variables->contains (k: decl))
2161 {
2162 HOST_WIDE_INT size = offsets[l - 3] - offsets[l - 2];
2163 if (dump_file && (dump_flags & TDF_DETAILS))
2164 {
2165 const char *n = (DECL_NAME (decl)
2166 ? IDENTIFIER_POINTER (DECL_NAME (decl))
2167 : "<unknown>");
2168 fprintf (stream: dump_file, format: "Unpoisoning shadow stack for variable: "
2169 "%s (%" PRId64 " B)\n", n, size);
2170 }
2171
2172 last_size += size & ~(ASAN_MIN_RED_ZONE_SIZE - HOST_WIDE_INT_1);
2173 }
2174 }
2175 last_size_aligned
2176 = ((last_size + (ASAN_RED_ZONE_SIZE - HOST_WIDE_INT_1))
2177 & ~(ASAN_RED_ZONE_SIZE - HOST_WIDE_INT_1));
2178 }
2179 if (last_size_aligned)
2180 {
2181 shadow_mem = adjust_address (shadow_mem, VOIDmode,
2182 (last_offset - prev_offset)
2183 >> ASAN_SHADOW_SHIFT);
2184 asan_clear_shadow (shadow_mem, len: last_size_aligned >> ASAN_SHADOW_SHIFT);
2185 }
2186
2187 /* Clean-up set with instrumented stack variables. */
2188 delete asan_handled_variables;
2189 asan_handled_variables = NULL;
2190 delete asan_used_labels;
2191 asan_used_labels = NULL;
2192
2193 do_pending_stack_adjust ();
2194 if (lab)
2195 emit_label (lab);
2196
2197 insns = get_insns ();
2198 end_sequence ();
2199 return insns;
2200}
2201
2202/* Emit __asan_allocas_unpoison (top, bot) call. The BASE parameter corresponds
2203 to BOT argument, for TOP virtual_stack_dynamic_rtx is used. NEW_SEQUENCE
2204 indicates whether we're emitting new instructions sequence or not. */
2205
2206rtx_insn *
2207asan_emit_allocas_unpoison (rtx top, rtx bot, rtx_insn *before)
2208{
2209 if (before)
2210 push_to_sequence (before);
2211 else
2212 start_sequence ();
2213 rtx ret = init_one_libfunc ("__asan_allocas_unpoison");
2214 top = convert_memory_address (ptr_mode, top);
2215 bot = convert_memory_address (ptr_mode, bot);
2216 emit_library_call (fun: ret, fn_type: LCT_NORMAL, outmode: ptr_mode,
2217 arg1: top, arg1_mode: ptr_mode, arg2: bot, arg2_mode: ptr_mode);
2218
2219 do_pending_stack_adjust ();
2220 rtx_insn *insns = get_insns ();
2221 end_sequence ();
2222 return insns;
2223}
2224
2225/* Return true if DECL, a global var, might be overridden and needs
2226 therefore a local alias. */
2227
2228static bool
2229asan_needs_local_alias (tree decl)
2230{
2231 return DECL_WEAK (decl) || !targetm.binds_local_p (decl);
2232}
2233
2234/* Return true if DECL, a global var, is an artificial ODR indicator symbol
2235 therefore doesn't need protection. */
2236
2237static bool
2238is_odr_indicator (tree decl)
2239{
2240 return (DECL_ARTIFICIAL (decl)
2241 && lookup_attribute (attr_name: "asan odr indicator", DECL_ATTRIBUTES (decl)));
2242}
2243
2244/* Return true if DECL is a VAR_DECL that should be protected
2245 by Address Sanitizer, by appending a red zone with protected
2246 shadow memory after it and aligning it to at least
2247 ASAN_RED_ZONE_SIZE bytes. */
2248
2249bool
2250asan_protect_global (tree decl, bool ignore_decl_rtl_set_p)
2251{
2252 if (!param_asan_globals)
2253 return false;
2254
2255 rtx rtl, symbol;
2256
2257 if (TREE_CODE (decl) == STRING_CST)
2258 {
2259 /* Instrument all STRING_CSTs except those created
2260 by asan_pp_string here. */
2261 if (shadow_ptr_types[0] != NULL_TREE
2262 && TREE_CODE (TREE_TYPE (decl)) == ARRAY_TYPE
2263 && TREE_TYPE (TREE_TYPE (decl)) == TREE_TYPE (shadow_ptr_types[0]))
2264 return false;
2265 return true;
2266 }
2267 if (!VAR_P (decl)
2268 /* TLS vars aren't statically protectable. */
2269 || DECL_THREAD_LOCAL_P (decl)
2270 /* Externs will be protected elsewhere. */
2271 || DECL_EXTERNAL (decl)
2272 /* PR sanitizer/81697: For architectures that use section anchors first
2273 call to asan_protect_global may occur before DECL_RTL (decl) is set.
2274 We should ignore DECL_RTL_SET_P then, because otherwise the first call
2275 to asan_protect_global will return FALSE and the following calls on the
2276 same decl after setting DECL_RTL (decl) will return TRUE and we'll end
2277 up with inconsistency at runtime. */
2278 || (!DECL_RTL_SET_P (decl) && !ignore_decl_rtl_set_p)
2279 /* Comdat vars pose an ABI problem, we can't know if
2280 the var that is selected by the linker will have
2281 padding or not. */
2282 || DECL_ONE_ONLY (decl)
2283 /* Similarly for common vars. People can use -fno-common.
2284 Note: Linux kernel is built with -fno-common, so we do instrument
2285 globals there even if it is C. */
2286 || (DECL_COMMON (decl) && TREE_PUBLIC (decl))
2287 /* Don't protect if using user section, often vars placed
2288 into user section from multiple TUs are then assumed
2289 to be an array of such vars, putting padding in there
2290 breaks this assumption. */
2291 || (DECL_SECTION_NAME (decl) != NULL
2292 && !symtab_node::get (decl)->implicit_section
2293 && !section_sanitized_p (DECL_SECTION_NAME (decl)))
2294 || DECL_SIZE (decl) == 0
2295 || ASAN_RED_ZONE_SIZE * BITS_PER_UNIT > MAX_OFILE_ALIGNMENT
2296 || TREE_CODE (DECL_SIZE_UNIT (decl)) != INTEGER_CST
2297 || !valid_constant_size_p (DECL_SIZE_UNIT (decl))
2298 || DECL_ALIGN_UNIT (decl) > 2 * ASAN_RED_ZONE_SIZE
2299 || TREE_TYPE (decl) == ubsan_get_source_location_type ()
2300 || is_odr_indicator (decl))
2301 return false;
2302
2303 if (!ignore_decl_rtl_set_p || DECL_RTL_SET_P (decl))
2304 {
2305
2306 rtl = DECL_RTL (decl);
2307 if (!MEM_P (rtl) || GET_CODE (XEXP (rtl, 0)) != SYMBOL_REF)
2308 return false;
2309 symbol = XEXP (rtl, 0);
2310
2311 if (CONSTANT_POOL_ADDRESS_P (symbol)
2312 || TREE_CONSTANT_POOL_ADDRESS_P (symbol))
2313 return false;
2314 }
2315
2316 if (lookup_attribute (attr_name: "weakref", DECL_ATTRIBUTES (decl)))
2317 return false;
2318
2319 if (!TARGET_SUPPORTS_ALIASES && asan_needs_local_alias (decl))
2320 return false;
2321
2322 return true;
2323}
2324
2325/* Construct a function tree for __asan_report_{load,store}{1,2,4,8,16,_n}.
2326 IS_STORE is either 1 (for a store) or 0 (for a load). */
2327
2328static tree
2329report_error_func (bool is_store, bool recover_p, HOST_WIDE_INT size_in_bytes,
2330 int *nargs)
2331{
2332 gcc_assert (!hwasan_sanitize_p ());
2333
2334 static enum built_in_function report[2][2][6]
2335 = { { { BUILT_IN_ASAN_REPORT_LOAD1, BUILT_IN_ASAN_REPORT_LOAD2,
2336 BUILT_IN_ASAN_REPORT_LOAD4, BUILT_IN_ASAN_REPORT_LOAD8,
2337 BUILT_IN_ASAN_REPORT_LOAD16, BUILT_IN_ASAN_REPORT_LOAD_N },
2338 { BUILT_IN_ASAN_REPORT_STORE1, BUILT_IN_ASAN_REPORT_STORE2,
2339 BUILT_IN_ASAN_REPORT_STORE4, BUILT_IN_ASAN_REPORT_STORE8,
2340 BUILT_IN_ASAN_REPORT_STORE16, BUILT_IN_ASAN_REPORT_STORE_N } },
2341 { { BUILT_IN_ASAN_REPORT_LOAD1_NOABORT,
2342 BUILT_IN_ASAN_REPORT_LOAD2_NOABORT,
2343 BUILT_IN_ASAN_REPORT_LOAD4_NOABORT,
2344 BUILT_IN_ASAN_REPORT_LOAD8_NOABORT,
2345 BUILT_IN_ASAN_REPORT_LOAD16_NOABORT,
2346 BUILT_IN_ASAN_REPORT_LOAD_N_NOABORT },
2347 { BUILT_IN_ASAN_REPORT_STORE1_NOABORT,
2348 BUILT_IN_ASAN_REPORT_STORE2_NOABORT,
2349 BUILT_IN_ASAN_REPORT_STORE4_NOABORT,
2350 BUILT_IN_ASAN_REPORT_STORE8_NOABORT,
2351 BUILT_IN_ASAN_REPORT_STORE16_NOABORT,
2352 BUILT_IN_ASAN_REPORT_STORE_N_NOABORT } } };
2353 if (size_in_bytes == -1)
2354 {
2355 *nargs = 2;
2356 return builtin_decl_implicit (fncode: report[recover_p][is_store][5]);
2357 }
2358 *nargs = 1;
2359 int size_log2 = exact_log2 (x: size_in_bytes);
2360 return builtin_decl_implicit (fncode: report[recover_p][is_store][size_log2]);
2361}
2362
2363/* Construct a function tree for __asan_{load,store}{1,2,4,8,16,_n}.
2364 IS_STORE is either 1 (for a store) or 0 (for a load). */
2365
2366static tree
2367check_func (bool is_store, bool recover_p, HOST_WIDE_INT size_in_bytes,
2368 int *nargs)
2369{
2370 static enum built_in_function check[2][2][6]
2371 = { { { BUILT_IN_ASAN_LOAD1, BUILT_IN_ASAN_LOAD2,
2372 BUILT_IN_ASAN_LOAD4, BUILT_IN_ASAN_LOAD8,
2373 BUILT_IN_ASAN_LOAD16, BUILT_IN_ASAN_LOADN },
2374 { BUILT_IN_ASAN_STORE1, BUILT_IN_ASAN_STORE2,
2375 BUILT_IN_ASAN_STORE4, BUILT_IN_ASAN_STORE8,
2376 BUILT_IN_ASAN_STORE16, BUILT_IN_ASAN_STOREN } },
2377 { { BUILT_IN_ASAN_LOAD1_NOABORT,
2378 BUILT_IN_ASAN_LOAD2_NOABORT,
2379 BUILT_IN_ASAN_LOAD4_NOABORT,
2380 BUILT_IN_ASAN_LOAD8_NOABORT,
2381 BUILT_IN_ASAN_LOAD16_NOABORT,
2382 BUILT_IN_ASAN_LOADN_NOABORT },
2383 { BUILT_IN_ASAN_STORE1_NOABORT,
2384 BUILT_IN_ASAN_STORE2_NOABORT,
2385 BUILT_IN_ASAN_STORE4_NOABORT,
2386 BUILT_IN_ASAN_STORE8_NOABORT,
2387 BUILT_IN_ASAN_STORE16_NOABORT,
2388 BUILT_IN_ASAN_STOREN_NOABORT } } };
2389 if (size_in_bytes == -1)
2390 {
2391 *nargs = 2;
2392 return builtin_decl_implicit (fncode: check[recover_p][is_store][5]);
2393 }
2394 *nargs = 1;
2395 int size_log2 = exact_log2 (x: size_in_bytes);
2396 return builtin_decl_implicit (fncode: check[recover_p][is_store][size_log2]);
2397}
2398
2399/* Split the current basic block and create a condition statement
2400 insertion point right before or after the statement pointed to by
2401 ITER. Return an iterator to the point at which the caller might
2402 safely insert the condition statement.
2403
2404 THEN_BLOCK must be set to the address of an uninitialized instance
2405 of basic_block. The function will then set *THEN_BLOCK to the
2406 'then block' of the condition statement to be inserted by the
2407 caller.
2408
2409 If CREATE_THEN_FALLTHRU_EDGE is false, no edge will be created from
2410 *THEN_BLOCK to *FALLTHROUGH_BLOCK.
2411
2412 Similarly, the function will set *FALLTRHOUGH_BLOCK to the 'else
2413 block' of the condition statement to be inserted by the caller.
2414
2415 Note that *FALLTHROUGH_BLOCK is a new block that contains the
2416 statements starting from *ITER, and *THEN_BLOCK is a new empty
2417 block.
2418
2419 *ITER is adjusted to point to always point to the first statement
2420 of the basic block * FALLTHROUGH_BLOCK. That statement is the
2421 same as what ITER was pointing to prior to calling this function,
2422 if BEFORE_P is true; otherwise, it is its following statement. */
2423
2424gimple_stmt_iterator
2425create_cond_insert_point (gimple_stmt_iterator *iter,
2426 bool before_p,
2427 bool then_more_likely_p,
2428 bool create_then_fallthru_edge,
2429 basic_block *then_block,
2430 basic_block *fallthrough_block)
2431{
2432 gimple_stmt_iterator gsi = *iter;
2433
2434 if (!gsi_end_p (i: gsi) && before_p)
2435 gsi_prev (i: &gsi);
2436
2437 basic_block cur_bb = gsi_bb (i: *iter);
2438
2439 edge e = split_block (cur_bb, gsi_stmt (i: gsi));
2440
2441 /* Get a hold on the 'condition block', the 'then block' and the
2442 'else block'. */
2443 basic_block cond_bb = e->src;
2444 basic_block fallthru_bb = e->dest;
2445 basic_block then_bb = create_empty_bb (cond_bb);
2446 if (current_loops)
2447 {
2448 add_bb_to_loop (then_bb, cond_bb->loop_father);
2449 loops_state_set (flags: LOOPS_NEED_FIXUP);
2450 }
2451
2452 /* Set up the newly created 'then block'. */
2453 e = make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE);
2454 profile_probability fallthrough_probability
2455 = then_more_likely_p
2456 ? profile_probability::very_unlikely ()
2457 : profile_probability::very_likely ();
2458 e->probability = fallthrough_probability.invert ();
2459 then_bb->count = e->count ();
2460 if (create_then_fallthru_edge)
2461 make_single_succ_edge (then_bb, fallthru_bb, EDGE_FALLTHRU);
2462
2463 /* Set up the fallthrough basic block. */
2464 e = find_edge (cond_bb, fallthru_bb);
2465 e->flags = EDGE_FALSE_VALUE;
2466 e->probability = fallthrough_probability;
2467
2468 /* Update dominance info for the newly created then_bb; note that
2469 fallthru_bb's dominance info has already been updated by
2470 split_bock. */
2471 if (dom_info_available_p (CDI_DOMINATORS))
2472 set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb);
2473
2474 *then_block = then_bb;
2475 *fallthrough_block = fallthru_bb;
2476 *iter = gsi_start_bb (bb: fallthru_bb);
2477
2478 return gsi_last_bb (bb: cond_bb);
2479}
2480
2481/* Insert an if condition followed by a 'then block' right before the
2482 statement pointed to by ITER. The fallthrough block -- which is the
2483 else block of the condition as well as the destination of the
2484 outcoming edge of the 'then block' -- starts with the statement
2485 pointed to by ITER.
2486
2487 COND is the condition of the if.
2488
2489 If THEN_MORE_LIKELY_P is true, the probability of the edge to the
2490 'then block' is higher than the probability of the edge to the
2491 fallthrough block.
2492
2493 Upon completion of the function, *THEN_BB is set to the newly
2494 inserted 'then block' and similarly, *FALLTHROUGH_BB is set to the
2495 fallthrough block.
2496
2497 *ITER is adjusted to still point to the same statement it was
2498 pointing to initially. */
2499
2500static void
2501insert_if_then_before_iter (gcond *cond,
2502 gimple_stmt_iterator *iter,
2503 bool then_more_likely_p,
2504 basic_block *then_bb,
2505 basic_block *fallthrough_bb)
2506{
2507 gimple_stmt_iterator cond_insert_point =
2508 create_cond_insert_point (iter,
2509 /*before_p=*/true,
2510 then_more_likely_p,
2511 /*create_then_fallthru_edge=*/true,
2512 then_block: then_bb,
2513 fallthrough_block: fallthrough_bb);
2514 gsi_insert_after (&cond_insert_point, cond, GSI_NEW_STMT);
2515}
2516
2517/* Build (base_addr >> ASAN_SHADOW_SHIFT) + asan_shadow_offset ().
2518 If RETURN_ADDRESS is set to true, return memory location instread
2519 of a value in the shadow memory. */
2520
2521static tree
2522build_shadow_mem_access (gimple_stmt_iterator *gsi, location_t location,
2523 tree base_addr, tree shadow_ptr_type,
2524 bool return_address = false)
2525{
2526 tree t, uintptr_type = TREE_TYPE (base_addr);
2527 tree shadow_type = TREE_TYPE (shadow_ptr_type);
2528 gimple *g;
2529
2530 t = build_int_cst (uintptr_type, ASAN_SHADOW_SHIFT);
2531 g = gimple_build_assign (make_ssa_name (var: uintptr_type), RSHIFT_EXPR,
2532 base_addr, t);
2533 gimple_set_location (g, location);
2534 gsi_insert_after (gsi, g, GSI_NEW_STMT);
2535
2536 t = build_int_cst (uintptr_type, asan_shadow_offset ());
2537 g = gimple_build_assign (make_ssa_name (var: uintptr_type), PLUS_EXPR,
2538 gimple_assign_lhs (gs: g), t);
2539 gimple_set_location (g, location);
2540 gsi_insert_after (gsi, g, GSI_NEW_STMT);
2541
2542 g = gimple_build_assign (make_ssa_name (var: shadow_ptr_type), NOP_EXPR,
2543 gimple_assign_lhs (gs: g));
2544 gimple_set_location (g, location);
2545 gsi_insert_after (gsi, g, GSI_NEW_STMT);
2546
2547 if (!return_address)
2548 {
2549 t = build2 (MEM_REF, shadow_type, gimple_assign_lhs (gs: g),
2550 build_int_cst (shadow_ptr_type, 0));
2551 g = gimple_build_assign (make_ssa_name (var: shadow_type), MEM_REF, t);
2552 gimple_set_location (g, location);
2553 gsi_insert_after (gsi, g, GSI_NEW_STMT);
2554 }
2555
2556 return gimple_assign_lhs (gs: g);
2557}
2558
2559/* BASE can already be an SSA_NAME; in that case, do not create a
2560 new SSA_NAME for it. */
2561
2562static tree
2563maybe_create_ssa_name (location_t loc, tree base, gimple_stmt_iterator *iter,
2564 bool before_p)
2565{
2566 STRIP_USELESS_TYPE_CONVERSION (base);
2567 if (TREE_CODE (base) == SSA_NAME)
2568 return base;
2569 gimple *g = gimple_build_assign (make_ssa_name (TREE_TYPE (base)), base);
2570 gimple_set_location (g, location: loc);
2571 if (before_p)
2572 gsi_insert_before (iter, g, GSI_SAME_STMT);
2573 else
2574 gsi_insert_after (iter, g, GSI_NEW_STMT);
2575 return gimple_assign_lhs (gs: g);
2576}
2577
2578/* LEN can already have necessary size and precision;
2579 in that case, do not create a new variable. */
2580
2581tree
2582maybe_cast_to_ptrmode (location_t loc, tree len, gimple_stmt_iterator *iter,
2583 bool before_p)
2584{
2585 if (ptrofftype_p (type: len))
2586 return len;
2587 gimple *g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
2588 NOP_EXPR, len);
2589 gimple_set_location (g, location: loc);
2590 if (before_p)
2591 gsi_insert_before (iter, g, GSI_SAME_STMT);
2592 else
2593 gsi_insert_after (iter, g, GSI_NEW_STMT);
2594 return gimple_assign_lhs (gs: g);
2595}
2596
2597/* Instrument the memory access instruction BASE. Insert new
2598 statements before or after ITER.
2599
2600 Note that the memory access represented by BASE can be either an
2601 SSA_NAME, or a non-SSA expression. LOCATION is the source code
2602 location. IS_STORE is TRUE for a store, FALSE for a load.
2603 BEFORE_P is TRUE for inserting the instrumentation code before
2604 ITER, FALSE for inserting it after ITER. IS_SCALAR_ACCESS is TRUE
2605 for a scalar memory access and FALSE for memory region access.
2606 NON_ZERO_P is TRUE if memory region is guaranteed to have non-zero
2607 length. ALIGN tells alignment of accessed memory object.
2608
2609 START_INSTRUMENTED and END_INSTRUMENTED are TRUE if start/end of
2610 memory region have already been instrumented.
2611
2612 If BEFORE_P is TRUE, *ITER is arranged to still point to the
2613 statement it was pointing to prior to calling this function,
2614 otherwise, it points to the statement logically following it. */
2615
2616static void
2617build_check_stmt (location_t loc, tree base, tree len,
2618 HOST_WIDE_INT size_in_bytes, gimple_stmt_iterator *iter,
2619 bool is_non_zero_len, bool before_p, bool is_store,
2620 bool is_scalar_access, unsigned int align = 0)
2621{
2622 gimple_stmt_iterator gsi = *iter;
2623 gimple *g;
2624
2625 gcc_assert (!(size_in_bytes > 0 && !is_non_zero_len));
2626 gcc_assert (size_in_bytes == -1 || size_in_bytes >= 1);
2627
2628 gsi = *iter;
2629
2630 base = unshare_expr (base);
2631 base = maybe_create_ssa_name (loc, base, iter: &gsi, before_p);
2632
2633 if (len)
2634 {
2635 len = unshare_expr (len);
2636 len = maybe_cast_to_ptrmode (loc, len, iter, before_p);
2637 }
2638 else
2639 {
2640 gcc_assert (size_in_bytes != -1);
2641 len = build_int_cst (pointer_sized_int_node, size_in_bytes);
2642 }
2643
2644 if (size_in_bytes > 1)
2645 {
2646 if ((size_in_bytes & (size_in_bytes - 1)) != 0
2647 || size_in_bytes > 16)
2648 is_scalar_access = false;
2649 else if (align && align < size_in_bytes * BITS_PER_UNIT)
2650 {
2651 /* On non-strict alignment targets, if
2652 16-byte access is just 8-byte aligned,
2653 this will result in misaligned shadow
2654 memory 2 byte load, but otherwise can
2655 be handled using one read. */
2656 if (size_in_bytes != 16
2657 || STRICT_ALIGNMENT
2658 || align < 8 * BITS_PER_UNIT)
2659 is_scalar_access = false;
2660 }
2661 }
2662
2663 HOST_WIDE_INT flags = 0;
2664 if (is_store)
2665 flags |= ASAN_CHECK_STORE;
2666 if (is_non_zero_len)
2667 flags |= ASAN_CHECK_NON_ZERO_LEN;
2668 if (is_scalar_access)
2669 flags |= ASAN_CHECK_SCALAR_ACCESS;
2670
2671 enum internal_fn fn = hwasan_sanitize_p ()
2672 ? IFN_HWASAN_CHECK
2673 : IFN_ASAN_CHECK;
2674
2675 g = gimple_build_call_internal (fn, 4,
2676 build_int_cst (integer_type_node, flags),
2677 base, len,
2678 build_int_cst (integer_type_node,
2679 align / BITS_PER_UNIT));
2680 gimple_set_location (g, location: loc);
2681 if (before_p)
2682 gsi_insert_before (&gsi, g, GSI_SAME_STMT);
2683 else
2684 {
2685 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
2686 gsi_next (i: &gsi);
2687 *iter = gsi;
2688 }
2689}
2690
2691/* If T represents a memory access, add instrumentation code before ITER.
2692 LOCATION is source code location.
2693 IS_STORE is either TRUE (for a store) or FALSE (for a load). */
2694
2695static void
2696instrument_derefs (gimple_stmt_iterator *iter, tree t,
2697 location_t location, bool is_store)
2698{
2699 if (is_store && !(asan_instrument_writes () || hwasan_instrument_writes ()))
2700 return;
2701 if (!is_store && !(asan_instrument_reads () || hwasan_instrument_reads ()))
2702 return;
2703
2704 tree type, base;
2705 HOST_WIDE_INT size_in_bytes;
2706 if (location == UNKNOWN_LOCATION)
2707 location = EXPR_LOCATION (t);
2708
2709 type = TREE_TYPE (t);
2710 switch (TREE_CODE (t))
2711 {
2712 case ARRAY_REF:
2713 case COMPONENT_REF:
2714 case INDIRECT_REF:
2715 case MEM_REF:
2716 case VAR_DECL:
2717 case BIT_FIELD_REF:
2718 break;
2719 /* FALLTHRU */
2720 default:
2721 return;
2722 }
2723
2724 size_in_bytes = int_size_in_bytes (type);
2725 if (size_in_bytes <= 0)
2726 return;
2727
2728 poly_int64 bitsize, bitpos;
2729 tree offset;
2730 machine_mode mode;
2731 int unsignedp, reversep, volatilep = 0;
2732 tree inner = get_inner_reference (t, &bitsize, &bitpos, &offset, &mode,
2733 &unsignedp, &reversep, &volatilep);
2734
2735 if (TREE_CODE (t) == COMPONENT_REF
2736 && DECL_BIT_FIELD_REPRESENTATIVE (TREE_OPERAND (t, 1)) != NULL_TREE)
2737 {
2738 tree repr = DECL_BIT_FIELD_REPRESENTATIVE (TREE_OPERAND (t, 1));
2739 instrument_derefs (iter, t: build3 (COMPONENT_REF, TREE_TYPE (repr),
2740 TREE_OPERAND (t, 0), repr,
2741 TREE_OPERAND (t, 2)),
2742 location, is_store);
2743 return;
2744 }
2745
2746 if (!multiple_p (a: bitpos, BITS_PER_UNIT)
2747 || maybe_ne (a: bitsize, b: size_in_bytes * BITS_PER_UNIT))
2748 return;
2749
2750 if (VAR_P (inner) && DECL_HARD_REGISTER (inner))
2751 return;
2752
2753 poly_int64 decl_size;
2754 if ((VAR_P (inner) || TREE_CODE (inner) == RESULT_DECL)
2755 && offset == NULL_TREE
2756 && DECL_SIZE (inner)
2757 && poly_int_tree_p (DECL_SIZE (inner), value: &decl_size)
2758 && known_subrange_p (pos1: bitpos, size1: bitsize, pos2: 0, size2: decl_size))
2759 {
2760 if (VAR_P (inner) && DECL_THREAD_LOCAL_P (inner))
2761 return;
2762 /* If we're not sanitizing globals and we can tell statically that this
2763 access is inside a global variable, then there's no point adding
2764 instrumentation to check the access. N.b. hwasan currently never
2765 sanitizes globals. */
2766 if ((hwasan_sanitize_p () || !param_asan_globals)
2767 && is_global_var (t: inner))
2768 return;
2769 if (!TREE_STATIC (inner))
2770 {
2771 /* Automatic vars in the current function will be always
2772 accessible. */
2773 if (decl_function_context (inner) == current_function_decl
2774 && (!asan_sanitize_use_after_scope ()
2775 || !TREE_ADDRESSABLE (inner)))
2776 return;
2777 }
2778 /* Always instrument external vars, they might be dynamically
2779 initialized. */
2780 else if (!DECL_EXTERNAL (inner))
2781 {
2782 /* For static vars if they are known not to be dynamically
2783 initialized, they will be always accessible. */
2784 varpool_node *vnode = varpool_node::get (decl: inner);
2785 if (vnode && !vnode->dynamically_initialized)
2786 return;
2787 }
2788 }
2789
2790 if (DECL_P (inner)
2791 && decl_function_context (inner) == current_function_decl
2792 && !TREE_ADDRESSABLE (inner))
2793 mark_addressable (inner);
2794
2795 base = build_fold_addr_expr (t);
2796 if (!has_mem_ref_been_instrumented (ref: base, access_size: size_in_bytes))
2797 {
2798 unsigned int align = get_object_alignment (t);
2799 build_check_stmt (loc: location, base, NULL_TREE, size_in_bytes, iter,
2800 /*is_non_zero_len*/size_in_bytes > 0, /*before_p=*/true,
2801 is_store, /*is_scalar_access*/true, align);
2802 update_mem_ref_hash_table (ref: base, access_size: size_in_bytes);
2803 update_mem_ref_hash_table (ref: t, access_size: size_in_bytes);
2804 }
2805
2806}
2807
2808/* Insert a memory reference into the hash table if access length
2809 can be determined in compile time. */
2810
2811static void
2812maybe_update_mem_ref_hash_table (tree base, tree len)
2813{
2814 if (!POINTER_TYPE_P (TREE_TYPE (base))
2815 || !INTEGRAL_TYPE_P (TREE_TYPE (len)))
2816 return;
2817
2818 HOST_WIDE_INT size_in_bytes = tree_fits_shwi_p (len) ? tree_to_shwi (len) : -1;
2819
2820 if (size_in_bytes != -1)
2821 update_mem_ref_hash_table (ref: base, access_size: size_in_bytes);
2822}
2823
2824/* Instrument an access to a contiguous memory region that starts at
2825 the address pointed to by BASE, over a length of LEN (expressed in
2826 the sizeof (*BASE) bytes). ITER points to the instruction before
2827 which the instrumentation instructions must be inserted. LOCATION
2828 is the source location that the instrumentation instructions must
2829 have. If IS_STORE is true, then the memory access is a store;
2830 otherwise, it's a load. */
2831
2832static void
2833instrument_mem_region_access (tree base, tree len,
2834 gimple_stmt_iterator *iter,
2835 location_t location, bool is_store)
2836{
2837 if (!POINTER_TYPE_P (TREE_TYPE (base))
2838 || !INTEGRAL_TYPE_P (TREE_TYPE (len))
2839 || integer_zerop (len))
2840 return;
2841
2842 HOST_WIDE_INT size_in_bytes = tree_fits_shwi_p (len) ? tree_to_shwi (len) : -1;
2843
2844 if ((size_in_bytes == -1)
2845 || !has_mem_ref_been_instrumented (ref: base, access_size: size_in_bytes))
2846 {
2847 build_check_stmt (loc: location, base, len, size_in_bytes, iter,
2848 /*is_non_zero_len*/size_in_bytes > 0, /*before_p*/true,
2849 is_store, /*is_scalar_access*/false, /*align*/0);
2850 }
2851
2852 maybe_update_mem_ref_hash_table (base, len);
2853 *iter = gsi_for_stmt (gsi_stmt (i: *iter));
2854}
2855
2856/* Instrument the call to a built-in memory access function that is
2857 pointed to by the iterator ITER.
2858
2859 Upon completion, return TRUE iff *ITER has been advanced to the
2860 statement following the one it was originally pointing to. */
2861
2862static bool
2863instrument_builtin_call (gimple_stmt_iterator *iter)
2864{
2865 if (!(asan_memintrin () || hwasan_memintrin ()))
2866 return false;
2867
2868 bool iter_advanced_p = false;
2869 gcall *call = as_a <gcall *> (p: gsi_stmt (i: *iter));
2870
2871 gcc_checking_assert (gimple_call_builtin_p (call, BUILT_IN_NORMAL));
2872
2873 location_t loc = gimple_location (g: call);
2874
2875 asan_mem_ref src0, src1, dest;
2876 asan_mem_ref_init (ref: &src0, NULL, access_size: 1);
2877 asan_mem_ref_init (ref: &src1, NULL, access_size: 1);
2878 asan_mem_ref_init (ref: &dest, NULL, access_size: 1);
2879
2880 tree src0_len = NULL_TREE, src1_len = NULL_TREE, dest_len = NULL_TREE;
2881 bool src0_is_store = false, src1_is_store = false, dest_is_store = false,
2882 dest_is_deref = false, intercepted_p = true;
2883
2884 if (get_mem_refs_of_builtin_call (call,
2885 src0: &src0, src0_len: &src0_len, src0_is_store: &src0_is_store,
2886 src1: &src1, src1_len: &src1_len, src1_is_store: &src1_is_store,
2887 dst: &dest, dst_len: &dest_len, dst_is_store: &dest_is_store,
2888 dest_is_deref: &dest_is_deref, intercepted_p: &intercepted_p, iter))
2889 {
2890 if (dest_is_deref)
2891 {
2892 instrument_derefs (iter, t: dest.start, location: loc, is_store: dest_is_store);
2893 gsi_next (i: iter);
2894 iter_advanced_p = true;
2895 }
2896 else if (!intercepted_p
2897 && (src0_len || src1_len || dest_len))
2898 {
2899 if (src0.start != NULL_TREE)
2900 instrument_mem_region_access (base: src0.start, len: src0_len,
2901 iter, location: loc, /*is_store=*/false);
2902 if (src1.start != NULL_TREE)
2903 instrument_mem_region_access (base: src1.start, len: src1_len,
2904 iter, location: loc, /*is_store=*/false);
2905 if (dest.start != NULL_TREE)
2906 instrument_mem_region_access (base: dest.start, len: dest_len,
2907 iter, location: loc, /*is_store=*/true);
2908
2909 *iter = gsi_for_stmt (call);
2910 gsi_next (i: iter);
2911 iter_advanced_p = true;
2912 }
2913 else
2914 {
2915 if (src0.start != NULL_TREE)
2916 maybe_update_mem_ref_hash_table (base: src0.start, len: src0_len);
2917 if (src1.start != NULL_TREE)
2918 maybe_update_mem_ref_hash_table (base: src1.start, len: src1_len);
2919 if (dest.start != NULL_TREE)
2920 maybe_update_mem_ref_hash_table (base: dest.start, len: dest_len);
2921 }
2922 }
2923 return iter_advanced_p;
2924}
2925
2926/* Instrument the assignment statement ITER if it is subject to
2927 instrumentation. Return TRUE iff instrumentation actually
2928 happened. In that case, the iterator ITER is advanced to the next
2929 logical expression following the one initially pointed to by ITER,
2930 and the relevant memory reference that which access has been
2931 instrumented is added to the memory references hash table. */
2932
2933static bool
2934maybe_instrument_assignment (gimple_stmt_iterator *iter)
2935{
2936 gimple *s = gsi_stmt (i: *iter);
2937
2938 gcc_assert (gimple_assign_single_p (s));
2939
2940 tree ref_expr = NULL_TREE;
2941 bool is_store, is_instrumented = false;
2942
2943 if (gimple_store_p (gs: s))
2944 {
2945 ref_expr = gimple_assign_lhs (gs: s);
2946 is_store = true;
2947 instrument_derefs (iter, t: ref_expr,
2948 location: gimple_location (g: s),
2949 is_store);
2950 is_instrumented = true;
2951 }
2952
2953 if (gimple_assign_load_p (s))
2954 {
2955 ref_expr = gimple_assign_rhs1 (gs: s);
2956 is_store = false;
2957 instrument_derefs (iter, t: ref_expr,
2958 location: gimple_location (g: s),
2959 is_store);
2960 is_instrumented = true;
2961 }
2962
2963 if (is_instrumented)
2964 gsi_next (i: iter);
2965
2966 return is_instrumented;
2967}
2968
2969/* Instrument the function call pointed to by the iterator ITER, if it
2970 is subject to instrumentation. At the moment, the only function
2971 calls that are instrumented are some built-in functions that access
2972 memory. Look at instrument_builtin_call to learn more.
2973
2974 Upon completion return TRUE iff *ITER was advanced to the statement
2975 following the one it was originally pointing to. */
2976
2977static bool
2978maybe_instrument_call (gimple_stmt_iterator *iter)
2979{
2980 gimple *stmt = gsi_stmt (i: *iter);
2981 bool is_builtin = gimple_call_builtin_p (stmt, BUILT_IN_NORMAL);
2982
2983 if (is_builtin && instrument_builtin_call (iter))
2984 return true;
2985
2986 if (gimple_call_noreturn_p (s: stmt))
2987 {
2988 if (is_builtin)
2989 {
2990 tree callee = gimple_call_fndecl (gs: stmt);
2991 switch (DECL_FUNCTION_CODE (decl: callee))
2992 {
2993 case BUILT_IN_UNREACHABLE:
2994 case BUILT_IN_UNREACHABLE_TRAP:
2995 case BUILT_IN_TRAP:
2996 /* Don't instrument these. */
2997 return false;
2998 default:
2999 break;
3000 }
3001 }
3002 /* If a function does not return, then we must handle clearing up the
3003 shadow stack accordingly. For ASAN we can simply set the entire stack
3004 to "valid" for accesses by setting the shadow space to 0 and all
3005 accesses will pass checks. That means that some bad accesses may be
3006 missed, but we will not report any false positives.
3007
3008 This is not possible for HWASAN. Since there is no "always valid" tag
3009 we can not set any space to "always valid". If we were to clear the
3010 entire shadow stack then code resuming from `longjmp` or a caught
3011 exception would trigger false positives when correctly accessing
3012 variables on the stack. Hence we need to handle things like
3013 `longjmp`, thread exit, and exceptions in a different way. These
3014 problems must be handled externally to the compiler, e.g. in the
3015 language runtime. */
3016 if (! hwasan_sanitize_p ())
3017 {
3018 tree decl = builtin_decl_implicit (fncode: BUILT_IN_ASAN_HANDLE_NO_RETURN);
3019 gimple *g = gimple_build_call (decl, 0);
3020 gimple_set_location (g, location: gimple_location (g: stmt));
3021 gsi_insert_before (iter, g, GSI_SAME_STMT);
3022 }
3023 }
3024
3025 bool instrumented = false;
3026 if (gimple_store_p (gs: stmt))
3027 {
3028 tree ref_expr = gimple_call_lhs (gs: stmt);
3029 instrument_derefs (iter, t: ref_expr,
3030 location: gimple_location (g: stmt),
3031 /*is_store=*/true);
3032
3033 instrumented = true;
3034 }
3035
3036 /* Walk through gimple_call arguments and check them id needed. */
3037 unsigned args_num = gimple_call_num_args (gs: stmt);
3038 for (unsigned i = 0; i < args_num; ++i)
3039 {
3040 tree arg = gimple_call_arg (gs: stmt, index: i);
3041 /* If ARG is not a non-aggregate register variable, compiler in general
3042 creates temporary for it and pass it as argument to gimple call.
3043 But in some cases, e.g. when we pass by value a small structure that
3044 fits to register, compiler can avoid extra overhead by pulling out
3045 these temporaries. In this case, we should check the argument. */
3046 if (!is_gimple_reg (arg) && !is_gimple_min_invariant (arg))
3047 {
3048 instrument_derefs (iter, t: arg,
3049 location: gimple_location (g: stmt),
3050 /*is_store=*/false);
3051 instrumented = true;
3052 }
3053 }
3054 if (instrumented)
3055 gsi_next (i: iter);
3056 return instrumented;
3057}
3058
3059/* Walk each instruction of all basic block and instrument those that
3060 represent memory references: loads, stores, or function calls.
3061 In a given basic block, this function avoids instrumenting memory
3062 references that have already been instrumented. */
3063
3064static void
3065transform_statements (void)
3066{
3067 basic_block bb, last_bb = NULL;
3068 gimple_stmt_iterator i;
3069 int saved_last_basic_block = last_basic_block_for_fn (cfun);
3070
3071 FOR_EACH_BB_FN (bb, cfun)
3072 {
3073 basic_block prev_bb = bb;
3074
3075 if (bb->index >= saved_last_basic_block) continue;
3076
3077 /* Flush the mem ref hash table, if current bb doesn't have
3078 exactly one predecessor, or if that predecessor (skipping
3079 over asan created basic blocks) isn't the last processed
3080 basic block. Thus we effectively flush on extended basic
3081 block boundaries. */
3082 while (single_pred_p (bb: prev_bb))
3083 {
3084 prev_bb = single_pred (bb: prev_bb);
3085 if (prev_bb->index < saved_last_basic_block)
3086 break;
3087 }
3088 if (prev_bb != last_bb)
3089 empty_mem_ref_hash_table ();
3090 last_bb = bb;
3091
3092 for (i = gsi_start_bb (bb); !gsi_end_p (i);)
3093 {
3094 gimple *s = gsi_stmt (i);
3095
3096 if (has_stmt_been_instrumented_p (stmt: s))
3097 gsi_next (i: &i);
3098 else if (gimple_assign_single_p (gs: s)
3099 && !gimple_clobber_p (s)
3100 && maybe_instrument_assignment (iter: &i))
3101 /* Nothing to do as maybe_instrument_assignment advanced
3102 the iterator I. */;
3103 else if (is_gimple_call (gs: s) && maybe_instrument_call (iter: &i))
3104 /* Nothing to do as maybe_instrument_call
3105 advanced the iterator I. */;
3106 else
3107 {
3108 /* No instrumentation happened.
3109
3110 If the current instruction is a function call that
3111 might free something, let's forget about the memory
3112 references that got instrumented. Otherwise we might
3113 miss some instrumentation opportunities. Do the same
3114 for a ASAN_MARK poisoning internal function. */
3115 if (is_gimple_call (gs: s)
3116 && (!nonfreeing_call_p (s)
3117 || asan_mark_p (stmt: s, flag: ASAN_MARK_POISON)))
3118 empty_mem_ref_hash_table ();
3119
3120 gsi_next (i: &i);
3121 }
3122 }
3123 }
3124 free_mem_ref_resources ();
3125}
3126
3127/* Build
3128 __asan_before_dynamic_init (module_name)
3129 or
3130 __asan_after_dynamic_init ()
3131 call. */
3132
3133tree
3134asan_dynamic_init_call (bool after_p)
3135{
3136 if (shadow_ptr_types[0] == NULL_TREE)
3137 asan_init_shadow_ptr_types ();
3138
3139 tree fn = builtin_decl_implicit (fncode: after_p
3140 ? BUILT_IN_ASAN_AFTER_DYNAMIC_INIT
3141 : BUILT_IN_ASAN_BEFORE_DYNAMIC_INIT);
3142 tree module_name_cst = NULL_TREE;
3143 if (!after_p)
3144 {
3145 pretty_printer module_name_pp;
3146 pp_string (&module_name_pp, main_input_filename);
3147
3148 module_name_cst = asan_pp_string (pp: &module_name_pp);
3149 module_name_cst = fold_convert (const_ptr_type_node,
3150 module_name_cst);
3151 }
3152
3153 return build_call_expr (fn, after_p ? 0 : 1, module_name_cst);
3154}
3155
3156/* Build
3157 struct __asan_global
3158 {
3159 const void *__beg;
3160 uptr __size;
3161 uptr __size_with_redzone;
3162 const void *__name;
3163 const void *__module_name;
3164 uptr __has_dynamic_init;
3165 __asan_global_source_location *__location;
3166 char *__odr_indicator;
3167 } type. */
3168
3169static tree
3170asan_global_struct (void)
3171{
3172 static const char *field_names[]
3173 = { "__beg", "__size", "__size_with_redzone",
3174 "__name", "__module_name", "__has_dynamic_init", "__location",
3175 "__odr_indicator" };
3176 tree fields[ARRAY_SIZE (field_names)], ret;
3177 unsigned i;
3178
3179 ret = make_node (RECORD_TYPE);
3180 for (i = 0; i < ARRAY_SIZE (field_names); i++)
3181 {
3182 fields[i]
3183 = build_decl (UNKNOWN_LOCATION, FIELD_DECL,
3184 get_identifier (field_names[i]),
3185 (i == 0 || i == 3) ? const_ptr_type_node
3186 : pointer_sized_int_node);
3187 DECL_CONTEXT (fields[i]) = ret;
3188 if (i)
3189 DECL_CHAIN (fields[i - 1]) = fields[i];
3190 }
3191 tree type_decl = build_decl (input_location, TYPE_DECL,
3192 get_identifier ("__asan_global"), ret);
3193 DECL_IGNORED_P (type_decl) = 1;
3194 DECL_ARTIFICIAL (type_decl) = 1;
3195 TYPE_FIELDS (ret) = fields[0];
3196 TYPE_NAME (ret) = type_decl;
3197 TYPE_STUB_DECL (ret) = type_decl;
3198 TYPE_ARTIFICIAL (ret) = 1;
3199 layout_type (ret);
3200 return ret;
3201}
3202
3203/* Create and return odr indicator symbol for DECL.
3204 TYPE is __asan_global struct type as returned by asan_global_struct. */
3205
3206static tree
3207create_odr_indicator (tree decl, tree type)
3208{
3209 char *name;
3210 tree uptr = TREE_TYPE (DECL_CHAIN (TYPE_FIELDS (type)));
3211 tree decl_name
3212 = (HAS_DECL_ASSEMBLER_NAME_P (decl) ? DECL_ASSEMBLER_NAME (decl)
3213 : DECL_NAME (decl));
3214 /* DECL_NAME theoretically might be NULL. Bail out with 0 in this case. */
3215 if (decl_name == NULL_TREE)
3216 return build_int_cst (uptr, 0);
3217 const char *dname = IDENTIFIER_POINTER (decl_name);
3218 if (HAS_DECL_ASSEMBLER_NAME_P (decl))
3219 dname = targetm.strip_name_encoding (dname);
3220 size_t len = strlen (s: dname) + sizeof ("__odr_asan_");
3221 name = XALLOCAVEC (char, len);
3222 snprintf (s: name, maxlen: len, format: "__odr_asan_%s", dname);
3223#ifndef NO_DOT_IN_LABEL
3224 name[sizeof ("__odr_asan") - 1] = '.';
3225#elif !defined(NO_DOLLAR_IN_LABEL)
3226 name[sizeof ("__odr_asan") - 1] = '$';
3227#endif
3228 tree var = build_decl (UNKNOWN_LOCATION, VAR_DECL, get_identifier (name),
3229 char_type_node);
3230 TREE_ADDRESSABLE (var) = 1;
3231 TREE_READONLY (var) = 0;
3232 TREE_THIS_VOLATILE (var) = 1;
3233 DECL_ARTIFICIAL (var) = 1;
3234 DECL_IGNORED_P (var) = 1;
3235 TREE_STATIC (var) = 1;
3236 TREE_PUBLIC (var) = 1;
3237 DECL_VISIBILITY (var) = DECL_VISIBILITY (decl);
3238 DECL_VISIBILITY_SPECIFIED (var) = DECL_VISIBILITY_SPECIFIED (decl);
3239
3240 TREE_USED (var) = 1;
3241 tree ctor = build_constructor_va (TREE_TYPE (var), 1, NULL_TREE,
3242 build_int_cst (unsigned_type_node, 0));
3243 TREE_CONSTANT (ctor) = 1;
3244 TREE_STATIC (ctor) = 1;
3245 DECL_INITIAL (var) = ctor;
3246 DECL_ATTRIBUTES (var) = tree_cons (get_identifier ("asan odr indicator"),
3247 NULL, DECL_ATTRIBUTES (var));
3248 make_decl_rtl (var);
3249 varpool_node::finalize_decl (decl: var);
3250 return fold_convert (uptr, build_fold_addr_expr (var));
3251}
3252
3253/* Return true if DECL, a global var, might be overridden and needs
3254 an additional odr indicator symbol. */
3255
3256static bool
3257asan_needs_odr_indicator_p (tree decl)
3258{
3259 /* Don't emit ODR indicators for kernel because:
3260 a) Kernel is written in C thus doesn't need ODR indicators.
3261 b) Some kernel code may have assumptions about symbols containing specific
3262 patterns in their names. Since ODR indicators contain original names
3263 of symbols they are emitted for, these assumptions would be broken for
3264 ODR indicator symbols. */
3265 return (!(flag_sanitize & SANITIZE_KERNEL_ADDRESS)
3266 && !DECL_ARTIFICIAL (decl)
3267 && !DECL_WEAK (decl)
3268 && TREE_PUBLIC (decl));
3269}
3270
3271/* Append description of a single global DECL into vector V.
3272 TYPE is __asan_global struct type as returned by asan_global_struct. */
3273
3274static void
3275asan_add_global (tree decl, tree type, vec<constructor_elt, va_gc> *v)
3276{
3277 tree init, uptr = TREE_TYPE (DECL_CHAIN (TYPE_FIELDS (type)));
3278 unsigned HOST_WIDE_INT size;
3279 tree str_cst, module_name_cst, refdecl = decl;
3280 vec<constructor_elt, va_gc> *vinner = NULL;
3281
3282 pretty_printer asan_pp, module_name_pp;
3283
3284 if (DECL_NAME (decl))
3285 pp_tree_identifier (&asan_pp, DECL_NAME (decl));
3286 else
3287 pp_string (&asan_pp, "<unknown>");
3288 str_cst = asan_pp_string (pp: &asan_pp);
3289
3290 if (!in_lto_p)
3291 pp_string (&module_name_pp, main_input_filename);
3292 else
3293 {
3294 const_tree tu = get_ultimate_context ((const_tree)decl);
3295 if (tu != NULL_TREE)
3296 pp_string (&module_name_pp, IDENTIFIER_POINTER (DECL_NAME (tu)));
3297 else
3298 pp_string (&module_name_pp, aux_base_name);
3299 }
3300
3301 module_name_cst = asan_pp_string (pp: &module_name_pp);
3302
3303 if (asan_needs_local_alias (decl))
3304 {
3305 char buf[20];
3306 ASM_GENERATE_INTERNAL_LABEL (buf, "LASAN", vec_safe_length (v) + 1);
3307 refdecl = build_decl (DECL_SOURCE_LOCATION (decl),
3308 VAR_DECL, get_identifier (buf), TREE_TYPE (decl));
3309 TREE_ADDRESSABLE (refdecl) = TREE_ADDRESSABLE (decl);
3310 TREE_READONLY (refdecl) = TREE_READONLY (decl);
3311 TREE_THIS_VOLATILE (refdecl) = TREE_THIS_VOLATILE (decl);
3312 DECL_NOT_GIMPLE_REG_P (refdecl) = DECL_NOT_GIMPLE_REG_P (decl);
3313 DECL_ARTIFICIAL (refdecl) = DECL_ARTIFICIAL (decl);
3314 DECL_IGNORED_P (refdecl) = DECL_IGNORED_P (decl);
3315 TREE_STATIC (refdecl) = 1;
3316 TREE_PUBLIC (refdecl) = 0;
3317 TREE_USED (refdecl) = 1;
3318 assemble_alias (refdecl, DECL_ASSEMBLER_NAME (decl));
3319 }
3320
3321 tree odr_indicator_ptr
3322 = (asan_needs_odr_indicator_p (decl) ? create_odr_indicator (decl, type)
3323 : build_int_cst (uptr, 0));
3324 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE,
3325 fold_convert (const_ptr_type_node,
3326 build_fold_addr_expr (refdecl)));
3327 size = tree_to_uhwi (DECL_SIZE_UNIT (decl));
3328 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE, build_int_cst (uptr, size));
3329 size += asan_red_zone_size (size);
3330 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE, build_int_cst (uptr, size));
3331 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE,
3332 fold_convert (const_ptr_type_node, str_cst));
3333 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE,
3334 fold_convert (const_ptr_type_node, module_name_cst));
3335 varpool_node *vnode = varpool_node::get (decl);
3336 int has_dynamic_init = 0;
3337 /* FIXME: Enable initialization order fiasco detection in LTO mode once
3338 proper fix for PR 79061 will be applied. */
3339 if (!in_lto_p)
3340 has_dynamic_init = vnode ? vnode->dynamically_initialized : 0;
3341 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE,
3342 build_int_cst (uptr, has_dynamic_init));
3343 tree locptr = NULL_TREE;
3344 location_t loc = DECL_SOURCE_LOCATION (decl);
3345 expanded_location xloc = expand_location (loc);
3346 if (xloc.file != NULL)
3347 {
3348 static int lasanloccnt = 0;
3349 char buf[25];
3350 ASM_GENERATE_INTERNAL_LABEL (buf, "LASANLOC", ++lasanloccnt);
3351 tree var = build_decl (UNKNOWN_LOCATION, VAR_DECL, get_identifier (buf),
3352 ubsan_get_source_location_type ());
3353 TREE_STATIC (var) = 1;
3354 TREE_PUBLIC (var) = 0;
3355 DECL_ARTIFICIAL (var) = 1;
3356 DECL_IGNORED_P (var) = 1;
3357 pretty_printer filename_pp;
3358 pp_string (&filename_pp, xloc.file);
3359 tree str = asan_pp_string (pp: &filename_pp);
3360 tree ctor = build_constructor_va (TREE_TYPE (var), 3,
3361 NULL_TREE, str, NULL_TREE,
3362 build_int_cst (unsigned_type_node,
3363 xloc.line), NULL_TREE,
3364 build_int_cst (unsigned_type_node,
3365 xloc.column));
3366 TREE_CONSTANT (ctor) = 1;
3367 TREE_STATIC (ctor) = 1;
3368 DECL_INITIAL (var) = ctor;
3369 varpool_node::finalize_decl (decl: var);
3370 locptr = fold_convert (uptr, build_fold_addr_expr (var));
3371 }
3372 else
3373 locptr = build_int_cst (uptr, 0);
3374 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE, locptr);
3375 CONSTRUCTOR_APPEND_ELT (vinner, NULL_TREE, odr_indicator_ptr);
3376 init = build_constructor (type, vinner);
3377 CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, init);
3378}
3379
3380/* Initialize sanitizer.def builtins if the FE hasn't initialized them. */
3381void
3382initialize_sanitizer_builtins (void)
3383{
3384 tree decl;
3385
3386 if (builtin_decl_implicit_p (fncode: BUILT_IN_ASAN_INIT))
3387 return;
3388
3389 tree BT_FN_VOID = build_function_type_list (void_type_node, NULL_TREE);
3390 tree BT_FN_VOID_PTR
3391 = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE);
3392 tree BT_FN_VOID_CONST_PTR
3393 = build_function_type_list (void_type_node, const_ptr_type_node, NULL_TREE);
3394 tree BT_FN_VOID_PTR_PTR
3395 = build_function_type_list (void_type_node, ptr_type_node,
3396 ptr_type_node, NULL_TREE);
3397 tree BT_FN_VOID_PTR_PTR_PTR
3398 = build_function_type_list (void_type_node, ptr_type_node,
3399 ptr_type_node, ptr_type_node, NULL_TREE);
3400 tree BT_FN_VOID_PTR_PTRMODE
3401 = build_function_type_list (void_type_node, ptr_type_node,
3402 pointer_sized_int_node, NULL_TREE);
3403 tree BT_FN_VOID_INT
3404 = build_function_type_list (void_type_node, integer_type_node, NULL_TREE);
3405 tree BT_FN_SIZE_CONST_PTR_INT
3406 = build_function_type_list (size_type_node, const_ptr_type_node,
3407 integer_type_node, NULL_TREE);
3408
3409 tree BT_FN_VOID_UINT8_UINT8
3410 = build_function_type_list (void_type_node, unsigned_char_type_node,
3411 unsigned_char_type_node, NULL_TREE);
3412 tree BT_FN_VOID_UINT16_UINT16
3413 = build_function_type_list (void_type_node, uint16_type_node,
3414 uint16_type_node, NULL_TREE);
3415 tree BT_FN_VOID_UINT32_UINT32
3416 = build_function_type_list (void_type_node, uint32_type_node,
3417 uint32_type_node, NULL_TREE);
3418 tree BT_FN_VOID_UINT64_UINT64
3419 = build_function_type_list (void_type_node, uint64_type_node,
3420 uint64_type_node, NULL_TREE);
3421 tree BT_FN_VOID_FLOAT_FLOAT
3422 = build_function_type_list (void_type_node, float_type_node,
3423 float_type_node, NULL_TREE);
3424 tree BT_FN_VOID_DOUBLE_DOUBLE
3425 = build_function_type_list (void_type_node, double_type_node,
3426 double_type_node, NULL_TREE);
3427 tree BT_FN_VOID_UINT64_PTR
3428 = build_function_type_list (void_type_node, uint64_type_node,
3429 ptr_type_node, NULL_TREE);
3430
3431 tree BT_FN_PTR_CONST_PTR_UINT8
3432 = build_function_type_list (ptr_type_node, const_ptr_type_node,
3433 unsigned_char_type_node, NULL_TREE);
3434 tree BT_FN_VOID_PTR_UINT8_PTRMODE
3435 = build_function_type_list (void_type_node, ptr_type_node,
3436 unsigned_char_type_node,
3437 pointer_sized_int_node, NULL_TREE);
3438
3439 tree BT_FN_BOOL_VPTR_PTR_IX_INT_INT[5];
3440 tree BT_FN_IX_CONST_VPTR_INT[5];
3441 tree BT_FN_IX_VPTR_IX_INT[5];
3442 tree BT_FN_VOID_VPTR_IX_INT[5];
3443 tree vptr
3444 = build_pointer_type (build_qualified_type (void_type_node,
3445 TYPE_QUAL_VOLATILE));
3446 tree cvptr
3447 = build_pointer_type (build_qualified_type (void_type_node,
3448 TYPE_QUAL_VOLATILE
3449 |TYPE_QUAL_CONST));
3450 tree boolt
3451 = lang_hooks.types.type_for_size (BOOL_TYPE_SIZE, 1);
3452 int i;
3453 for (i = 0; i < 5; i++)
3454 {
3455 tree ix = build_nonstandard_integer_type (BITS_PER_UNIT * (1 << i), 1);
3456 BT_FN_BOOL_VPTR_PTR_IX_INT_INT[i]
3457 = build_function_type_list (boolt, vptr, ptr_type_node, ix,
3458 integer_type_node, integer_type_node,
3459 NULL_TREE);
3460 BT_FN_IX_CONST_VPTR_INT[i]
3461 = build_function_type_list (ix, cvptr, integer_type_node, NULL_TREE);
3462 BT_FN_IX_VPTR_IX_INT[i]
3463 = build_function_type_list (ix, vptr, ix, integer_type_node,
3464 NULL_TREE);
3465 BT_FN_VOID_VPTR_IX_INT[i]
3466 = build_function_type_list (void_type_node, vptr, ix,
3467 integer_type_node, NULL_TREE);
3468 }
3469#define BT_FN_BOOL_VPTR_PTR_I1_INT_INT BT_FN_BOOL_VPTR_PTR_IX_INT_INT[0]
3470#define BT_FN_I1_CONST_VPTR_INT BT_FN_IX_CONST_VPTR_INT[0]
3471#define BT_FN_I1_VPTR_I1_INT BT_FN_IX_VPTR_IX_INT[0]
3472#define BT_FN_VOID_VPTR_I1_INT BT_FN_VOID_VPTR_IX_INT[0]
3473#define BT_FN_BOOL_VPTR_PTR_I2_INT_INT BT_FN_BOOL_VPTR_PTR_IX_INT_INT[1]
3474#define BT_FN_I2_CONST_VPTR_INT BT_FN_IX_CONST_VPTR_INT[1]
3475#define BT_FN_I2_VPTR_I2_INT BT_FN_IX_VPTR_IX_INT[1]
3476#define BT_FN_VOID_VPTR_I2_INT BT_FN_VOID_VPTR_IX_INT[1]
3477#define BT_FN_BOOL_VPTR_PTR_I4_INT_INT BT_FN_BOOL_VPTR_PTR_IX_INT_INT[2]
3478#define BT_FN_I4_CONST_VPTR_INT BT_FN_IX_CONST_VPTR_INT[2]
3479#define BT_FN_I4_VPTR_I4_INT BT_FN_IX_VPTR_IX_INT[2]
3480#define BT_FN_VOID_VPTR_I4_INT BT_FN_VOID_VPTR_IX_INT[2]
3481#define BT_FN_BOOL_VPTR_PTR_I8_INT_INT BT_FN_BOOL_VPTR_PTR_IX_INT_INT[3]
3482#define BT_FN_I8_CONST_VPTR_INT BT_FN_IX_CONST_VPTR_INT[3]
3483#define BT_FN_I8_VPTR_I8_INT BT_FN_IX_VPTR_IX_INT[3]
3484#define BT_FN_VOID_VPTR_I8_INT BT_FN_VOID_VPTR_IX_INT[3]
3485#define BT_FN_BOOL_VPTR_PTR_I16_INT_INT BT_FN_BOOL_VPTR_PTR_IX_INT_INT[4]
3486#define BT_FN_I16_CONST_VPTR_INT BT_FN_IX_CONST_VPTR_INT[4]
3487#define BT_FN_I16_VPTR_I16_INT BT_FN_IX_VPTR_IX_INT[4]
3488#define BT_FN_VOID_VPTR_I16_INT BT_FN_VOID_VPTR_IX_INT[4]
3489#undef ATTR_NOTHROW_LIST
3490#define ATTR_NOTHROW_LIST ECF_NOTHROW
3491#undef ATTR_NOTHROW_LEAF_LIST
3492#define ATTR_NOTHROW_LEAF_LIST ECF_NOTHROW | ECF_LEAF
3493#undef ATTR_TMPURE_NOTHROW_LEAF_LIST
3494#define ATTR_TMPURE_NOTHROW_LEAF_LIST ECF_TM_PURE | ATTR_NOTHROW_LEAF_LIST
3495#undef ATTR_NORETURN_NOTHROW_LEAF_LIST
3496#define ATTR_NORETURN_NOTHROW_LEAF_LIST ECF_NORETURN | ATTR_NOTHROW_LEAF_LIST
3497#undef ATTR_CONST_NORETURN_NOTHROW_LEAF_LIST
3498#define ATTR_CONST_NORETURN_NOTHROW_LEAF_LIST \
3499 ECF_CONST | ATTR_NORETURN_NOTHROW_LEAF_LIST
3500#undef ATTR_TMPURE_NORETURN_NOTHROW_LEAF_LIST
3501#define ATTR_TMPURE_NORETURN_NOTHROW_LEAF_LIST \
3502 ECF_TM_PURE | ATTR_NORETURN_NOTHROW_LEAF_LIST
3503#undef ATTR_COLD_NOTHROW_LEAF_LIST
3504#define ATTR_COLD_NOTHROW_LEAF_LIST \
3505 /* ECF_COLD missing */ ATTR_NOTHROW_LEAF_LIST
3506#undef ATTR_COLD_NORETURN_NOTHROW_LEAF_LIST
3507#define ATTR_COLD_NORETURN_NOTHROW_LEAF_LIST \
3508 /* ECF_COLD missing */ ATTR_NORETURN_NOTHROW_LEAF_LIST
3509#undef ATTR_COLD_CONST_NORETURN_NOTHROW_LEAF_LIST
3510#define ATTR_COLD_CONST_NORETURN_NOTHROW_LEAF_LIST \
3511 /* ECF_COLD missing */ ATTR_CONST_NORETURN_NOTHROW_LEAF_LIST
3512#undef ATTR_PURE_NOTHROW_LEAF_LIST
3513#define ATTR_PURE_NOTHROW_LEAF_LIST ECF_PURE | ATTR_NOTHROW_LEAF_LIST
3514#undef DEF_BUILTIN_STUB
3515#define DEF_BUILTIN_STUB(ENUM, NAME)
3516#undef DEF_SANITIZER_BUILTIN_1
3517#define DEF_SANITIZER_BUILTIN_1(ENUM, NAME, TYPE, ATTRS) \
3518 do { \
3519 decl = add_builtin_function ("__builtin_" NAME, TYPE, ENUM, \
3520 BUILT_IN_NORMAL, NAME, NULL_TREE); \
3521 set_call_expr_flags (decl, ATTRS); \
3522 set_builtin_decl (ENUM, decl, true); \
3523 } while (0)
3524#undef DEF_SANITIZER_BUILTIN
3525#define DEF_SANITIZER_BUILTIN(ENUM, NAME, TYPE, ATTRS) \
3526 DEF_SANITIZER_BUILTIN_1 (ENUM, NAME, TYPE, ATTRS);
3527
3528#include "sanitizer.def"
3529
3530 /* -fsanitize=object-size uses __builtin_dynamic_object_size and
3531 __builtin_object_size, but they might not be available for e.g. Fortran at
3532 this point. We use DEF_SANITIZER_BUILTIN here only as a convenience
3533 macro. */
3534 if (flag_sanitize & SANITIZE_OBJECT_SIZE)
3535 {
3536 if (!builtin_decl_implicit_p (fncode: BUILT_IN_OBJECT_SIZE))
3537 DEF_SANITIZER_BUILTIN_1 (BUILT_IN_OBJECT_SIZE, "object_size",
3538 BT_FN_SIZE_CONST_PTR_INT,
3539 ATTR_PURE_NOTHROW_LEAF_LIST);
3540 if (!builtin_decl_implicit_p (fncode: BUILT_IN_DYNAMIC_OBJECT_SIZE))
3541 DEF_SANITIZER_BUILTIN_1 (BUILT_IN_DYNAMIC_OBJECT_SIZE,
3542 "dynamic_object_size",
3543 BT_FN_SIZE_CONST_PTR_INT,
3544 ATTR_PURE_NOTHROW_LEAF_LIST);
3545 }
3546
3547#undef DEF_SANITIZER_BUILTIN_1
3548#undef DEF_SANITIZER_BUILTIN
3549#undef DEF_BUILTIN_STUB
3550}
3551
3552/* Called via htab_traverse. Count number of emitted
3553 STRING_CSTs in the constant hash table. */
3554
3555int
3556count_string_csts (constant_descriptor_tree **slot,
3557 unsigned HOST_WIDE_INT *data)
3558{
3559 struct constant_descriptor_tree *desc = *slot;
3560 if (TREE_CODE (desc->value) == STRING_CST
3561 && TREE_ASM_WRITTEN (desc->value)
3562 && asan_protect_global (decl: desc->value))
3563 ++*data;
3564 return 1;
3565}
3566
3567/* Helper structure to pass two parameters to
3568 add_string_csts. */
3569
3570struct asan_add_string_csts_data
3571{
3572 tree type;
3573 vec<constructor_elt, va_gc> *v;
3574};
3575
3576/* Called via hash_table::traverse. Call asan_add_global
3577 on emitted STRING_CSTs from the constant hash table. */
3578
3579int
3580add_string_csts (constant_descriptor_tree **slot,
3581 asan_add_string_csts_data *aascd)
3582{
3583 struct constant_descriptor_tree *desc = *slot;
3584 if (TREE_CODE (desc->value) == STRING_CST
3585 && TREE_ASM_WRITTEN (desc->value)
3586 && asan_protect_global (decl: desc->value))
3587 {
3588 asan_add_global (SYMBOL_REF_DECL (XEXP (desc->rtl, 0)),
3589 type: aascd->type, v: aascd->v);
3590 }
3591 return 1;
3592}
3593
3594/* Needs to be GTY(()), because cgraph_build_static_cdtor may
3595 invoke ggc_collect. */
3596static GTY(()) tree asan_ctor_statements;
3597
3598/* Module-level instrumentation.
3599 - Insert __asan_init_vN() into the list of CTORs.
3600 - TODO: insert redzones around globals.
3601 */
3602
3603void
3604asan_finish_file (void)
3605{
3606 varpool_node *vnode;
3607 unsigned HOST_WIDE_INT gcount = 0;
3608
3609 if (shadow_ptr_types[0] == NULL_TREE)
3610 asan_init_shadow_ptr_types ();
3611 /* Avoid instrumenting code in the asan ctors/dtors.
3612 We don't need to insert padding after the description strings,
3613 nor after .LASAN* array. */
3614 flag_sanitize &= ~SANITIZE_ADDRESS;
3615
3616 /* For user-space we want asan constructors to run first.
3617 Linux kernel does not support priorities other than default, and the only
3618 other user of constructors is coverage. So we run with the default
3619 priority. */
3620 int priority = flag_sanitize & SANITIZE_USER_ADDRESS
3621 ? MAX_RESERVED_INIT_PRIORITY - 1 : DEFAULT_INIT_PRIORITY;
3622
3623 if (flag_sanitize & SANITIZE_USER_ADDRESS)
3624 {
3625 tree fn = builtin_decl_implicit (fncode: BUILT_IN_ASAN_INIT);
3626 append_to_statement_list (build_call_expr (fn, 0), &asan_ctor_statements);
3627 fn = builtin_decl_implicit (fncode: BUILT_IN_ASAN_VERSION_MISMATCH_CHECK);
3628 append_to_statement_list (build_call_expr (fn, 0), &asan_ctor_statements);
3629 }
3630 FOR_EACH_DEFINED_VARIABLE (vnode)
3631 if (TREE_ASM_WRITTEN (vnode->decl)
3632 && asan_protect_global (decl: vnode->decl))
3633 ++gcount;
3634 hash_table<tree_descriptor_hasher> *const_desc_htab = constant_pool_htab ();
3635 const_desc_htab->traverse<unsigned HOST_WIDE_INT *, count_string_csts>
3636 (argument: &gcount);
3637 if (gcount)
3638 {
3639 tree type = asan_global_struct (), var, ctor;
3640 tree dtor_statements = NULL_TREE;
3641 vec<constructor_elt, va_gc> *v;
3642 char buf[20];
3643
3644 type = build_array_type_nelts (type, gcount);
3645 ASM_GENERATE_INTERNAL_LABEL (buf, "LASAN", 0);
3646 var = build_decl (UNKNOWN_LOCATION, VAR_DECL, get_identifier (buf),
3647 type);
3648 TREE_STATIC (var) = 1;
3649 TREE_PUBLIC (var) = 0;
3650 DECL_ARTIFICIAL (var) = 1;
3651 DECL_IGNORED_P (var) = 1;
3652 vec_alloc (v, nelems: gcount);
3653 FOR_EACH_DEFINED_VARIABLE (vnode)
3654 if (TREE_ASM_WRITTEN (vnode->decl)
3655 && asan_protect_global (decl: vnode->decl))
3656 asan_add_global (decl: vnode->decl, TREE_TYPE (type), v);
3657 struct asan_add_string_csts_data aascd;
3658 aascd.type = TREE_TYPE (type);
3659 aascd.v = v;
3660 const_desc_htab->traverse<asan_add_string_csts_data *, add_string_csts>
3661 (argument: &aascd);
3662 ctor = build_constructor (type, v);
3663 TREE_CONSTANT (ctor) = 1;
3664 TREE_STATIC (ctor) = 1;
3665 DECL_INITIAL (var) = ctor;
3666 SET_DECL_ALIGN (var, MAX (DECL_ALIGN (var),
3667 ASAN_SHADOW_GRANULARITY * BITS_PER_UNIT));
3668
3669 varpool_node::finalize_decl (decl: var);
3670
3671 tree fn = builtin_decl_implicit (fncode: BUILT_IN_ASAN_REGISTER_GLOBALS);
3672 tree gcount_tree = build_int_cst (pointer_sized_int_node, gcount);
3673 append_to_statement_list (build_call_expr (fn, 2,
3674 build_fold_addr_expr (var),
3675 gcount_tree),
3676 &asan_ctor_statements);
3677
3678 fn = builtin_decl_implicit (fncode: BUILT_IN_ASAN_UNREGISTER_GLOBALS);
3679 append_to_statement_list (build_call_expr (fn, 2,
3680 build_fold_addr_expr (var),
3681 gcount_tree),
3682 &dtor_statements);
3683 cgraph_build_static_cdtor (which: 'D', body: dtor_statements, priority);
3684 }
3685 if (asan_ctor_statements)
3686 cgraph_build_static_cdtor (which: 'I', body: asan_ctor_statements, priority);
3687 flag_sanitize |= SANITIZE_ADDRESS;
3688}
3689
3690/* Poison or unpoison (depending on IS_CLOBBER variable) shadow memory based
3691 on SHADOW address. Newly added statements will be added to ITER with
3692 given location LOC. We mark SIZE bytes in shadow memory, where
3693 LAST_CHUNK_SIZE is greater than zero in situation where we are at the
3694 end of a variable. */
3695
3696static void
3697asan_store_shadow_bytes (gimple_stmt_iterator *iter, location_t loc,
3698 tree shadow,
3699 unsigned HOST_WIDE_INT base_addr_offset,
3700 bool is_clobber, unsigned size,
3701 unsigned last_chunk_size)
3702{
3703 tree shadow_ptr_type;
3704
3705 switch (size)
3706 {
3707 case 1:
3708 shadow_ptr_type = shadow_ptr_types[0];
3709 break;
3710 case 2:
3711 shadow_ptr_type = shadow_ptr_types[1];
3712 break;
3713 case 4:
3714 shadow_ptr_type = shadow_ptr_types[2];
3715 break;
3716 default:
3717 gcc_unreachable ();
3718 }
3719
3720 unsigned char c = (char) is_clobber ? ASAN_STACK_MAGIC_USE_AFTER_SCOPE : 0;
3721 unsigned HOST_WIDE_INT val = 0;
3722 unsigned last_pos = size;
3723 if (last_chunk_size && !is_clobber)
3724 last_pos = BYTES_BIG_ENDIAN ? 0 : size - 1;
3725 for (unsigned i = 0; i < size; ++i)
3726 {
3727 unsigned char shadow_c = c;
3728 if (i == last_pos)
3729 shadow_c = last_chunk_size;
3730 val |= (unsigned HOST_WIDE_INT) shadow_c << (BITS_PER_UNIT * i);
3731 }
3732
3733 /* Handle last chunk in unpoisoning. */
3734 tree magic = build_int_cst (TREE_TYPE (shadow_ptr_type), val);
3735
3736 tree dest = build2 (MEM_REF, TREE_TYPE (shadow_ptr_type), shadow,
3737 build_int_cst (shadow_ptr_type, base_addr_offset));
3738
3739 gimple *g = gimple_build_assign (dest, magic);
3740 gimple_set_location (g, location: loc);
3741 gsi_insert_after (iter, g, GSI_NEW_STMT);
3742}
3743
3744/* Expand the ASAN_MARK builtins. */
3745
3746bool
3747asan_expand_mark_ifn (gimple_stmt_iterator *iter)
3748{
3749 gimple *g = gsi_stmt (i: *iter);
3750 location_t loc = gimple_location (g);
3751 HOST_WIDE_INT flag = tree_to_shwi (gimple_call_arg (gs: g, index: 0));
3752 bool is_poison = ((asan_mark_flags)flag) == ASAN_MARK_POISON;
3753
3754 tree base = gimple_call_arg (gs: g, index: 1);
3755 gcc_checking_assert (TREE_CODE (base) == ADDR_EXPR);
3756 tree decl = TREE_OPERAND (base, 0);
3757
3758 /* For a nested function, we can have: ASAN_MARK (2, &FRAME.2.fp_input, 4) */
3759 if (TREE_CODE (decl) == COMPONENT_REF
3760 && DECL_NONLOCAL_FRAME (TREE_OPERAND (decl, 0)))
3761 decl = TREE_OPERAND (decl, 0);
3762
3763 gcc_checking_assert (TREE_CODE (decl) == VAR_DECL);
3764
3765 if (hwasan_sanitize_p ())
3766 {
3767 gcc_assert (param_hwasan_instrument_stack);
3768 gimple_seq stmts = NULL;
3769 /* Here we swap ASAN_MARK calls for HWASAN_MARK.
3770 This is because we are using the approach of using ASAN_MARK as a
3771 synonym until here.
3772 That approach means we don't yet have to duplicate all the special
3773 cases for ASAN_MARK and ASAN_POISON with the exact same handling but
3774 called HWASAN_MARK etc.
3775
3776 N.b. __asan_poison_stack_memory (which implements ASAN_MARK for ASAN)
3777 rounds the size up to its shadow memory granularity, while
3778 __hwasan_tag_memory (which implements the same for HWASAN) does not.
3779 Hence we emit HWASAN_MARK with an aligned size unlike ASAN_MARK. */
3780 tree len = gimple_call_arg (gs: g, index: 2);
3781 tree new_len = gimple_build_round_up (seq: &stmts, loc, size_type_node, old_size: len,
3782 HWASAN_TAG_GRANULE_SIZE);
3783 gimple_build (seq: &stmts, loc, fn: CFN_HWASAN_MARK,
3784 void_type_node, args: gimple_call_arg (gs: g, index: 0),
3785 args: base, args: new_len);
3786 gsi_replace_with_seq (iter, stmts, true);
3787 return false;
3788 }
3789
3790 if (is_poison)
3791 {
3792 if (asan_handled_variables == NULL)
3793 asan_handled_variables = new hash_set<tree> (16);
3794 asan_handled_variables->add (k: decl);
3795 }
3796 tree len = gimple_call_arg (gs: g, index: 2);
3797
3798 gcc_assert (tree_fits_shwi_p (len));
3799 unsigned HOST_WIDE_INT size_in_bytes = tree_to_shwi (len);
3800 gcc_assert (size_in_bytes);
3801
3802 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
3803 NOP_EXPR, base);
3804 gimple_set_location (g, location: loc);
3805 gsi_replace (iter, g, false);
3806 tree base_addr = gimple_assign_lhs (gs: g);
3807
3808 /* Generate direct emission if size_in_bytes is small. */
3809 if (size_in_bytes
3810 <= (unsigned)param_use_after_scope_direct_emission_threshold)
3811 {
3812 const unsigned HOST_WIDE_INT shadow_size
3813 = shadow_mem_size (size: size_in_bytes);
3814 const unsigned int shadow_align
3815 = (get_pointer_alignment (base) / BITS_PER_UNIT) >> ASAN_SHADOW_SHIFT;
3816
3817 tree shadow = build_shadow_mem_access (gsi: iter, location: loc, base_addr,
3818 shadow_ptr_type: shadow_ptr_types[0], return_address: true);
3819
3820 for (unsigned HOST_WIDE_INT offset = 0; offset < shadow_size;)
3821 {
3822 unsigned size = 1;
3823 if (shadow_size - offset >= 4
3824 && (!STRICT_ALIGNMENT || shadow_align >= 4))
3825 size = 4;
3826 else if (shadow_size - offset >= 2
3827 && (!STRICT_ALIGNMENT || shadow_align >= 2))
3828 size = 2;
3829
3830 unsigned HOST_WIDE_INT last_chunk_size = 0;
3831 unsigned HOST_WIDE_INT s = (offset + size) * ASAN_SHADOW_GRANULARITY;
3832 if (s > size_in_bytes)
3833 last_chunk_size = ASAN_SHADOW_GRANULARITY - (s - size_in_bytes);
3834
3835 asan_store_shadow_bytes (iter, loc, shadow, base_addr_offset: offset, is_clobber: is_poison,
3836 size, last_chunk_size);
3837 offset += size;
3838 }
3839 }
3840 else
3841 {
3842 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
3843 NOP_EXPR, len);
3844 gimple_set_location (g, location: loc);
3845 gsi_insert_before (iter, g, GSI_SAME_STMT);
3846 tree sz_arg = gimple_assign_lhs (gs: g);
3847
3848 tree fun
3849 = builtin_decl_implicit (fncode: is_poison ? BUILT_IN_ASAN_POISON_STACK_MEMORY
3850 : BUILT_IN_ASAN_UNPOISON_STACK_MEMORY);
3851 g = gimple_build_call (fun, 2, base_addr, sz_arg);
3852 gimple_set_location (g, location: loc);
3853 gsi_insert_after (iter, g, GSI_NEW_STMT);
3854 }
3855
3856 return false;
3857}
3858
3859/* Expand the ASAN_{LOAD,STORE} builtins. */
3860
3861bool
3862asan_expand_check_ifn (gimple_stmt_iterator *iter, bool use_calls)
3863{
3864 gcc_assert (!hwasan_sanitize_p ());
3865 gimple *g = gsi_stmt (i: *iter);
3866 location_t loc = gimple_location (g);
3867 bool recover_p;
3868 if (flag_sanitize & SANITIZE_USER_ADDRESS)
3869 recover_p = (flag_sanitize_recover & SANITIZE_USER_ADDRESS) != 0;
3870 else
3871 recover_p = (flag_sanitize_recover & SANITIZE_KERNEL_ADDRESS) != 0;
3872
3873 HOST_WIDE_INT flags = tree_to_shwi (gimple_call_arg (gs: g, index: 0));
3874 gcc_assert (flags < ASAN_CHECK_LAST);
3875 bool is_scalar_access = (flags & ASAN_CHECK_SCALAR_ACCESS) != 0;
3876 bool is_store = (flags & ASAN_CHECK_STORE) != 0;
3877 bool is_non_zero_len = (flags & ASAN_CHECK_NON_ZERO_LEN) != 0;
3878
3879 tree base = gimple_call_arg (gs: g, index: 1);
3880 tree len = gimple_call_arg (gs: g, index: 2);
3881 HOST_WIDE_INT align = tree_to_shwi (gimple_call_arg (gs: g, index: 3));
3882
3883 HOST_WIDE_INT size_in_bytes
3884 = is_scalar_access && tree_fits_shwi_p (len) ? tree_to_shwi (len) : -1;
3885
3886 if (use_calls)
3887 {
3888 /* Instrument using callbacks. */
3889 gimple *g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
3890 NOP_EXPR, base);
3891 gimple_set_location (g, location: loc);
3892 gsi_insert_before (iter, g, GSI_SAME_STMT);
3893 tree base_addr = gimple_assign_lhs (gs: g);
3894
3895 int nargs;
3896 tree fun = check_func (is_store, recover_p, size_in_bytes, nargs: &nargs);
3897 if (nargs == 1)
3898 g = gimple_build_call (fun, 1, base_addr);
3899 else
3900 {
3901 gcc_assert (nargs == 2);
3902 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
3903 NOP_EXPR, len);
3904 gimple_set_location (g, location: loc);
3905 gsi_insert_before (iter, g, GSI_SAME_STMT);
3906 tree sz_arg = gimple_assign_lhs (gs: g);
3907 g = gimple_build_call (fun, nargs, base_addr, sz_arg);
3908 }
3909 gimple_set_location (g, location: loc);
3910 gsi_replace (iter, g, false);
3911 return false;
3912 }
3913
3914 HOST_WIDE_INT real_size_in_bytes = size_in_bytes == -1 ? 1 : size_in_bytes;
3915
3916 tree shadow_ptr_type = shadow_ptr_types[real_size_in_bytes == 16 ? 1 : 0];
3917 tree shadow_type = TREE_TYPE (shadow_ptr_type);
3918
3919 gimple_stmt_iterator gsi = *iter;
3920
3921 if (!is_non_zero_len)
3922 {
3923 /* So, the length of the memory area to asan-protect is
3924 non-constant. Let's guard the generated instrumentation code
3925 like:
3926
3927 if (len != 0)
3928 {
3929 //asan instrumentation code goes here.
3930 }
3931 // falltrough instructions, starting with *ITER. */
3932
3933 g = gimple_build_cond (NE_EXPR,
3934 len,
3935 build_int_cst (TREE_TYPE (len), 0),
3936 NULL_TREE, NULL_TREE);
3937 gimple_set_location (g, location: loc);
3938
3939 basic_block then_bb, fallthrough_bb;
3940 insert_if_then_before_iter (cond: as_a <gcond *> (p: g), iter,
3941 /*then_more_likely_p=*/true,
3942 then_bb: &then_bb, fallthrough_bb: &fallthrough_bb);
3943 /* Note that fallthrough_bb starts with the statement that was
3944 pointed to by ITER. */
3945
3946 /* The 'then block' of the 'if (len != 0) condition is where
3947 we'll generate the asan instrumentation code now. */
3948 gsi = gsi_last_bb (bb: then_bb);
3949 }
3950
3951 /* Get an iterator on the point where we can add the condition
3952 statement for the instrumentation. */
3953 basic_block then_bb, else_bb;
3954 gsi = create_cond_insert_point (iter: &gsi, /*before_p*/false,
3955 /*then_more_likely_p=*/false,
3956 /*create_then_fallthru_edge*/recover_p,
3957 then_block: &then_bb,
3958 fallthrough_block: &else_bb);
3959
3960 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
3961 NOP_EXPR, base);
3962 gimple_set_location (g, location: loc);
3963 gsi_insert_before (&gsi, g, GSI_NEW_STMT);
3964 tree base_addr = gimple_assign_lhs (gs: g);
3965
3966 tree t = NULL_TREE;
3967 if (real_size_in_bytes >= 8)
3968 {
3969 tree shadow = build_shadow_mem_access (gsi: &gsi, location: loc, base_addr,
3970 shadow_ptr_type);
3971 t = shadow;
3972 }
3973 else
3974 {
3975 /* Slow path for 1, 2 and 4 byte accesses. */
3976 /* Test (shadow != 0)
3977 & ((base_addr & 7) + (real_size_in_bytes - 1)) >= shadow). */
3978 tree shadow = build_shadow_mem_access (gsi: &gsi, location: loc, base_addr,
3979 shadow_ptr_type);
3980 gimple *shadow_test = build_assign (NE_EXPR, shadow, 0);
3981 gimple_seq seq = NULL;
3982 gimple_seq_add_stmt (&seq, shadow_test);
3983 /* Aligned (>= 8 bytes) can test just
3984 (real_size_in_bytes - 1 >= shadow), as base_addr & 7 is known
3985 to be 0. */
3986 if (align < 8)
3987 {
3988 gimple_seq_add_stmt (&seq, build_assign (BIT_AND_EXPR,
3989 base_addr, 7));
3990 gimple_seq_add_stmt (&seq,
3991 build_type_cast (shadow_type,
3992 gimple_seq_last (s: seq)));
3993 if (real_size_in_bytes > 1)
3994 gimple_seq_add_stmt (&seq,
3995 build_assign (PLUS_EXPR,
3996 gimple_seq_last (s: seq),
3997 real_size_in_bytes - 1));
3998 t = gimple_assign_lhs (gs: gimple_seq_last_stmt (s: seq));
3999 }
4000 else
4001 t = build_int_cst (shadow_type, real_size_in_bytes - 1);
4002 gimple_seq_add_stmt (&seq, build_assign (GE_EXPR, t, shadow));
4003 gimple_seq_add_stmt (&seq, build_assign (BIT_AND_EXPR, shadow_test,
4004 gimple_seq_last (s: seq)));
4005 t = gimple_assign_lhs (gs: gimple_seq_last (s: seq));
4006 gimple_seq_set_location (seq, loc);
4007 gsi_insert_seq_after (&gsi, seq, GSI_CONTINUE_LINKING);
4008
4009 /* For non-constant, misaligned or otherwise weird access sizes,
4010 check first and last byte. */
4011 if (size_in_bytes == -1)
4012 {
4013 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
4014 MINUS_EXPR, len,
4015 build_int_cst (pointer_sized_int_node, 1));
4016 gimple_set_location (g, location: loc);
4017 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
4018 tree last = gimple_assign_lhs (gs: g);
4019 g = gimple_build_assign (make_ssa_name (pointer_sized_int_node),
4020 PLUS_EXPR, base_addr, last);
4021 gimple_set_location (g, location: loc);
4022 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
4023 tree base_end_addr = gimple_assign_lhs (gs: g);
4024
4025 tree shadow = build_shadow_mem_access (gsi: &gsi, location: loc, base_addr: base_end_addr,
4026 shadow_ptr_type);
4027 gimple *shadow_test = build_assign (NE_EXPR, shadow, 0);
4028 gimple_seq seq = NULL;
4029 gimple_seq_add_stmt (&seq, shadow_test);
4030 gimple_seq_add_stmt (&seq, build_assign (BIT_AND_EXPR,
4031 base_end_addr, 7));
4032 gimple_seq_add_stmt (&seq, build_type_cast (shadow_type,
4033 gimple_seq_last (s: seq)));
4034 gimple_seq_add_stmt (&seq, build_assign (GE_EXPR,
4035 gimple_seq_last (s: seq),
4036 shadow));
4037 gimple_seq_add_stmt (&seq, build_assign (BIT_AND_EXPR, shadow_test,
4038 gimple_seq_last (s: seq)));
4039 gimple_seq_add_stmt (&seq, build_assign (BIT_IOR_EXPR, t,
4040 gimple_seq_last (s: seq)));
4041 t = gimple_assign_lhs (gs: gimple_seq_last (s: seq));
4042 gimple_seq_set_location (seq, loc);
4043 gsi_insert_seq_after (&gsi, seq, GSI_CONTINUE_LINKING);
4044 }
4045 }
4046
4047 g = gimple_build_cond (NE_EXPR, t, build_int_cst (TREE_TYPE (t), 0),
4048 NULL_TREE, NULL_TREE);
4049 gimple_set_location (g, location: loc);
4050 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
4051
4052 /* Generate call to the run-time library (e.g. __asan_report_load8). */
4053 gsi = gsi_start_bb (bb: then_bb);
4054 int nargs;
4055 tree fun = report_error_func (is_store, recover_p, size_in_bytes, nargs: &nargs);
4056 g = gimple_build_call (fun, nargs, base_addr, len);
4057 gimple_set_location (g, location: loc);
4058 gsi_insert_after (&gsi, g, GSI_NEW_STMT);
4059
4060 gsi_remove (iter, true);
4061 *iter = gsi_start_bb (bb: else_bb);
4062
4063 return true;
4064}
4065
4066/* Create ASAN shadow variable for a VAR_DECL which has been rewritten
4067 into SSA. Already seen VAR_DECLs are stored in SHADOW_VARS_MAPPING. */
4068
4069static tree
4070create_asan_shadow_var (tree var_decl,
4071 hash_map<tree, tree> &shadow_vars_mapping)
4072{
4073 tree *slot = shadow_vars_mapping.get (k: var_decl);
4074 if (slot == NULL)
4075 {
4076 tree shadow_var = copy_node (var_decl);
4077
4078 copy_body_data id;
4079 memset (s: &id, c: 0, n: sizeof (copy_body_data));
4080 id.src_fn = id.dst_fn = current_function_decl;
4081 copy_decl_for_dup_finish (id: &id, decl: var_decl, copy: shadow_var);
4082
4083 DECL_ARTIFICIAL (shadow_var) = 1;
4084 DECL_IGNORED_P (shadow_var) = 1;
4085 DECL_SEEN_IN_BIND_EXPR_P (shadow_var) = 0;
4086 gimple_add_tmp_var (shadow_var);
4087
4088 shadow_vars_mapping.put (k: var_decl, v: shadow_var);
4089 return shadow_var;
4090 }
4091 else
4092 return *slot;
4093}
4094
4095/* Expand ASAN_POISON ifn. */
4096
4097bool
4098asan_expand_poison_ifn (gimple_stmt_iterator *iter,
4099 bool *need_commit_edge_insert,
4100 hash_map<tree, tree> &shadow_vars_mapping)
4101{
4102 gimple *g = gsi_stmt (i: *iter);
4103 tree poisoned_var = gimple_call_lhs (gs: g);
4104 if (!poisoned_var || has_zero_uses (var: poisoned_var))
4105 {
4106 gsi_remove (iter, true);
4107 return true;
4108 }
4109
4110 if (SSA_NAME_VAR (poisoned_var) == NULL_TREE)
4111 SET_SSA_NAME_VAR_OR_IDENTIFIER (poisoned_var,
4112 create_tmp_var (TREE_TYPE (poisoned_var)));
4113
4114 tree shadow_var = create_asan_shadow_var (SSA_NAME_VAR (poisoned_var),
4115 shadow_vars_mapping);
4116
4117 bool recover_p;
4118 if (flag_sanitize & SANITIZE_USER_ADDRESS)
4119 recover_p = (flag_sanitize_recover & SANITIZE_USER_ADDRESS) != 0;
4120 else
4121 recover_p = (flag_sanitize_recover & SANITIZE_KERNEL_ADDRESS) != 0;
4122 tree size = DECL_SIZE_UNIT (shadow_var);
4123 gimple *poison_call
4124 = gimple_build_call_internal (IFN_ASAN_MARK, 3,
4125 build_int_cst (integer_type_node,
4126 ASAN_MARK_POISON),
4127 build_fold_addr_expr (shadow_var), size);
4128
4129 gimple *use;
4130 imm_use_iterator imm_iter;
4131 FOR_EACH_IMM_USE_STMT (use, imm_iter, poisoned_var)
4132 {
4133 if (is_gimple_debug (gs: use))
4134 continue;
4135
4136 int nargs;
4137 bool store_p = gimple_call_internal_p (gs: use, fn: IFN_ASAN_POISON_USE);
4138 gcall *call;
4139 if (hwasan_sanitize_p ())
4140 {
4141 tree fun = builtin_decl_implicit (fncode: BUILT_IN_HWASAN_TAG_MISMATCH4);
4142 /* NOTE: hwasan has no __hwasan_report_* functions like asan does.
4143 We use __hwasan_tag_mismatch4 with arguments that tell it the
4144 size of access and load to report all tag mismatches.
4145
4146 The arguments to this function are:
4147 Address of invalid access.
4148 Bitfield containing information about the access
4149 (access_info)
4150 Pointer to a frame of registers
4151 (for use in printing the contents of registers in a dump)
4152 Not used yet -- to be used by inline instrumentation.
4153 Size of access.
4154
4155 The access_info bitfield encodes the following pieces of
4156 information:
4157 - Is this a store or load?
4158 access_info & 0x10 => store
4159 - Should the program continue after reporting the error?
4160 access_info & 0x20 => recover
4161 - What size access is this (not used here since we can always
4162 pass the size in the last argument)
4163
4164 if (access_info & 0xf == 0xf)
4165 size is taken from last argument.
4166 else
4167 size == 1 << (access_info & 0xf)
4168
4169 The last argument contains the size of the access iff the
4170 access_info size indicator is 0xf (we always use this argument
4171 rather than storing the size in the access_info bitfield).
4172
4173 See the function definition `__hwasan_tag_mismatch4` in
4174 libsanitizer/hwasan for the full definition.
4175 */
4176 unsigned access_info = (0x20 * recover_p)
4177 + (0x10 * store_p)
4178 + (0xf);
4179 call = gimple_build_call (fun, 4,
4180 build_fold_addr_expr (shadow_var),
4181 build_int_cst (pointer_sized_int_node,
4182 access_info),
4183 build_int_cst (pointer_sized_int_node, 0),
4184 size);
4185 }
4186 else
4187 {
4188 tree fun = report_error_func (is_store: store_p, recover_p, size_in_bytes: tree_to_uhwi (size),
4189 nargs: &nargs);
4190 call = gimple_build_call (fun, 1,
4191 build_fold_addr_expr (shadow_var));
4192 }
4193 gimple_set_location (g: call, location: gimple_location (g: use));
4194 gimple *call_to_insert = call;
4195
4196 /* The USE can be a gimple PHI node. If so, insert the call on
4197 all edges leading to the PHI node. */
4198 if (is_a <gphi *> (p: use))
4199 {
4200 gphi *phi = dyn_cast<gphi *> (p: use);
4201 for (unsigned i = 0; i < gimple_phi_num_args (gs: phi); ++i)
4202 if (gimple_phi_arg_def (gs: phi, index: i) == poisoned_var)
4203 {
4204 edge e = gimple_phi_arg_edge (phi, i);
4205
4206 /* Do not insert on an edge we can't split. */
4207 if (e->flags & EDGE_ABNORMAL)
4208 continue;
4209
4210 if (call_to_insert == NULL)
4211 call_to_insert = gimple_copy (call);
4212
4213 gsi_insert_seq_on_edge (e, call_to_insert);
4214 *need_commit_edge_insert = true;
4215 call_to_insert = NULL;
4216 }
4217 }
4218 else
4219 {
4220 gimple_stmt_iterator gsi = gsi_for_stmt (use);
4221 if (store_p)
4222 gsi_replace (&gsi, call, true);
4223 else
4224 gsi_insert_before (&gsi, call, GSI_NEW_STMT);
4225 }
4226 }
4227
4228 SSA_NAME_IS_DEFAULT_DEF (poisoned_var) = true;
4229 SSA_NAME_DEF_STMT (poisoned_var) = gimple_build_nop ();
4230 gsi_replace (iter, poison_call, false);
4231
4232 return true;
4233}
4234
4235/* Instrument the current function. */
4236
4237static unsigned int
4238asan_instrument (void)
4239{
4240 if (hwasan_sanitize_p ())
4241 {
4242 transform_statements ();
4243 return 0;
4244 }
4245
4246 if (shadow_ptr_types[0] == NULL_TREE)
4247 asan_init_shadow_ptr_types ();
4248 transform_statements ();
4249 last_alloca_addr = NULL_TREE;
4250 return 0;
4251}
4252
4253static bool
4254gate_asan (void)
4255{
4256 return sanitize_flags_p (flag: SANITIZE_ADDRESS);
4257}
4258
4259namespace {
4260
4261const pass_data pass_data_asan =
4262{
4263 .type: GIMPLE_PASS, /* type */
4264 .name: "asan", /* name */
4265 .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */
4266 .tv_id: TV_NONE, /* tv_id */
4267 .properties_required: ( PROP_ssa | PROP_cfg | PROP_gimple_leh ), /* properties_required */
4268 .properties_provided: 0, /* properties_provided */
4269 .properties_destroyed: 0, /* properties_destroyed */
4270 .todo_flags_start: 0, /* todo_flags_start */
4271 TODO_update_ssa, /* todo_flags_finish */
4272};
4273
4274class pass_asan : public gimple_opt_pass
4275{
4276public:
4277 pass_asan (gcc::context *ctxt)
4278 : gimple_opt_pass (pass_data_asan, ctxt)
4279 {}
4280
4281 /* opt_pass methods: */
4282 opt_pass * clone () final override { return new pass_asan (m_ctxt); }
4283 bool gate (function *) final override
4284 {
4285 return gate_asan () || gate_hwasan ();
4286 }
4287 unsigned int execute (function *) final override
4288 {
4289 return asan_instrument ();
4290 }
4291
4292}; // class pass_asan
4293
4294} // anon namespace
4295
4296gimple_opt_pass *
4297make_pass_asan (gcc::context *ctxt)
4298{
4299 return new pass_asan (ctxt);
4300}
4301
4302namespace {
4303
4304const pass_data pass_data_asan_O0 =
4305{
4306 .type: GIMPLE_PASS, /* type */
4307 .name: "asan0", /* name */
4308 .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */
4309 .tv_id: TV_NONE, /* tv_id */
4310 .properties_required: ( PROP_ssa | PROP_cfg | PROP_gimple_leh ), /* properties_required */
4311 .properties_provided: 0, /* properties_provided */
4312 .properties_destroyed: 0, /* properties_destroyed */
4313 .todo_flags_start: 0, /* todo_flags_start */
4314 TODO_update_ssa, /* todo_flags_finish */
4315};
4316
4317class pass_asan_O0 : public gimple_opt_pass
4318{
4319public:
4320 pass_asan_O0 (gcc::context *ctxt)
4321 : gimple_opt_pass (pass_data_asan_O0, ctxt)
4322 {}
4323
4324 /* opt_pass methods: */
4325 bool gate (function *) final override
4326 {
4327 return !optimize && (gate_asan () || gate_hwasan ());
4328 }
4329 unsigned int execute (function *) final override
4330 {
4331 return asan_instrument ();
4332 }
4333
4334}; // class pass_asan_O0
4335
4336} // anon namespace
4337
4338gimple_opt_pass *
4339make_pass_asan_O0 (gcc::context *ctxt)
4340{
4341 return new pass_asan_O0 (ctxt);
4342}
4343
4344/* HWASAN */
4345
4346/* For stack tagging:
4347
4348 Return the offset from the frame base tag that the "next" expanded object
4349 should have. */
4350uint8_t
4351hwasan_current_frame_tag ()
4352{
4353 return hwasan_frame_tag_offset;
4354}
4355
4356/* For stack tagging:
4357
4358 Return the 'base pointer' for this function. If that base pointer has not
4359 yet been created then we create a register to hold it and record the insns
4360 to initialize the register in `hwasan_frame_base_init_seq` for later
4361 emission. */
4362rtx
4363hwasan_frame_base ()
4364{
4365 if (! hwasan_frame_base_ptr)
4366 {
4367 start_sequence ();
4368 hwasan_frame_base_ptr
4369 = force_reg (Pmode,
4370 targetm.memtag.insert_random_tag (virtual_stack_vars_rtx,
4371 NULL_RTX));
4372 hwasan_frame_base_init_seq = get_insns ();
4373 end_sequence ();
4374 }
4375
4376 return hwasan_frame_base_ptr;
4377}
4378
4379/* For stack tagging:
4380
4381 Check whether this RTX is a standard pointer addressing the base of the
4382 stack variables for this frame. Returns true if the RTX is either
4383 virtual_stack_vars_rtx or hwasan_frame_base_ptr. */
4384bool
4385stack_vars_base_reg_p (rtx base)
4386{
4387 return base == virtual_stack_vars_rtx || base == hwasan_frame_base_ptr;
4388}
4389
4390/* For stack tagging:
4391
4392 Emit frame base initialisation.
4393 If hwasan_frame_base has been used before here then
4394 hwasan_frame_base_init_seq contains the sequence of instructions to
4395 initialize it. This must be put just before the hwasan prologue, so we emit
4396 the insns before parm_birth_insn (which will point to the first instruction
4397 of the hwasan prologue if it exists).
4398
4399 We update `parm_birth_insn` to point to the start of this initialisation
4400 since that represents the end of the initialisation done by
4401 expand_function_{start,end} functions and we want to maintain that. */
4402void
4403hwasan_maybe_emit_frame_base_init ()
4404{
4405 if (! hwasan_frame_base_init_seq)
4406 return;
4407 emit_insn_before (hwasan_frame_base_init_seq, parm_birth_insn);
4408 parm_birth_insn = hwasan_frame_base_init_seq;
4409}
4410
4411/* Record a compile-time constant size stack variable that HWASAN will need to
4412 tag. This record of the range of a stack variable will be used by
4413 `hwasan_emit_prologue` to emit the RTL at the start of each frame which will
4414 set tags in the shadow memory according to the assigned tag for each object.
4415
4416 The range that the object spans in stack space should be described by the
4417 bounds `untagged_base + nearest_offset` and
4418 `untagged_base + farthest_offset`.
4419 `tagged_base` is the base address which contains the "base frame tag" for
4420 this frame, and from which the value to address this object with will be
4421 calculated.
4422
4423 We record the `untagged_base` since the functions in the hwasan library we
4424 use to tag memory take pointers without a tag. */
4425void
4426hwasan_record_stack_var (rtx untagged_base, rtx tagged_base,
4427 poly_int64 nearest_offset, poly_int64 farthest_offset)
4428{
4429 hwasan_stack_var cur_var;
4430 cur_var.untagged_base = untagged_base;
4431 cur_var.tagged_base = tagged_base;
4432 cur_var.nearest_offset = nearest_offset;
4433 cur_var.farthest_offset = farthest_offset;
4434 cur_var.tag_offset = hwasan_current_frame_tag ();
4435
4436 hwasan_tagged_stack_vars.safe_push (obj: cur_var);
4437}
4438
4439/* Return the RTX representing the farthest extent of the statically allocated
4440 stack objects for this frame. If hwasan_frame_base_ptr has not been
4441 initialized then we are not storing any static variables on the stack in
4442 this frame. In this case we return NULL_RTX to represent that.
4443
4444 Otherwise simply return virtual_stack_vars_rtx + frame_offset. */
4445rtx
4446hwasan_get_frame_extent ()
4447{
4448 return (hwasan_frame_base_ptr
4449 ? plus_constant (Pmode, virtual_stack_vars_rtx, frame_offset)
4450 : NULL_RTX);
4451}
4452
4453/* For stack tagging:
4454
4455 Increment the frame tag offset modulo the size a tag can represent. */
4456void
4457hwasan_increment_frame_tag ()
4458{
4459 uint8_t tag_bits = HWASAN_TAG_SIZE;
4460 gcc_assert (HWASAN_TAG_SIZE
4461 <= sizeof (hwasan_frame_tag_offset) * CHAR_BIT);
4462 hwasan_frame_tag_offset = (hwasan_frame_tag_offset + 1) % (1 << tag_bits);
4463 /* The "background tag" of the stack is zero by definition.
4464 This is the tag that objects like parameters passed on the stack and
4465 spilled registers are given. It is handy to avoid this tag for objects
4466 whose tags we decide ourselves, partly to ensure that buffer overruns
4467 can't affect these important variables (e.g. saved link register, saved
4468 stack pointer etc) and partly to make debugging easier (everything with a
4469 tag of zero is space allocated automatically by the compiler).
4470
4471 This is not feasible when using random frame tags (the default
4472 configuration for hwasan) since the tag for the given frame is randomly
4473 chosen at runtime. In order to avoid any tags matching the stack
4474 background we would need to decide tag offsets at runtime instead of
4475 compile time (and pay the resulting performance cost).
4476
4477 When not using random base tags for each frame (i.e. when compiled with
4478 `--param hwasan-random-frame-tag=0`) the base tag for each frame is zero.
4479 This means the tag that each object gets is equal to the
4480 hwasan_frame_tag_offset used in determining it.
4481 When this is the case we *can* ensure no object gets the tag of zero by
4482 simply ensuring no object has the hwasan_frame_tag_offset of zero.
4483
4484 There is the extra complication that we only record the
4485 hwasan_frame_tag_offset here (which is the offset from the tag stored in
4486 the stack pointer). In the kernel, the tag in the stack pointer is 0xff
4487 rather than zero. This does not cause problems since tags of 0xff are
4488 never checked in the kernel. As mentioned at the beginning of this
4489 comment the background tag of the stack is zero by definition, which means
4490 that for the kernel we should skip offsets of both 0 and 1 from the stack
4491 pointer. Avoiding the offset of 0 ensures we use a tag which will be
4492 checked, avoiding the offset of 1 ensures we use a tag that is not the
4493 same as the background. */
4494 if (hwasan_frame_tag_offset == 0 && ! param_hwasan_random_frame_tag)
4495 hwasan_frame_tag_offset += 1;
4496 if (hwasan_frame_tag_offset == 1 && ! param_hwasan_random_frame_tag
4497 && sanitize_flags_p (flag: SANITIZE_KERNEL_HWADDRESS))
4498 hwasan_frame_tag_offset += 1;
4499}
4500
4501/* Clear internal state for the next function.
4502 This function is called before variables on the stack get expanded, in
4503 `init_vars_expansion`. */
4504void
4505hwasan_record_frame_init ()
4506{
4507 delete asan_used_labels;
4508 asan_used_labels = NULL;
4509
4510 /* If this isn't the case then some stack variable was recorded *before*
4511 hwasan_record_frame_init is called, yet *after* the hwasan prologue for
4512 the previous frame was emitted. Such stack variables would not have
4513 their shadow stack filled in. */
4514 gcc_assert (hwasan_tagged_stack_vars.is_empty ());
4515 hwasan_frame_base_ptr = NULL_RTX;
4516 hwasan_frame_base_init_seq = NULL;
4517
4518 /* When not using a random frame tag we can avoid the background stack
4519 color which gives the user a little better debug output upon a crash.
4520 Meanwhile, when using a random frame tag it will be nice to avoid adding
4521 tags for the first object since that is unnecessary extra work.
4522 Hence set the initial hwasan_frame_tag_offset to be 0 if using a random
4523 frame tag and 1 otherwise.
4524
4525 As described in hwasan_increment_frame_tag, in the kernel the stack
4526 pointer has the tag 0xff. That means that to avoid 0xff and 0 (the tag
4527 which the kernel does not check and the background tag respectively) we
4528 start with a tag offset of 2. */
4529 hwasan_frame_tag_offset = param_hwasan_random_frame_tag
4530 ? 0
4531 : sanitize_flags_p (flag: SANITIZE_KERNEL_HWADDRESS) ? 2 : 1;
4532}
4533
4534/* For stack tagging:
4535 (Emits HWASAN equivalent of what is emitted by
4536 `asan_emit_stack_protection`).
4537
4538 Emits the extra prologue code to set the shadow stack as required for HWASAN
4539 stack instrumentation.
4540
4541 Uses the vector of recorded stack variables hwasan_tagged_stack_vars. When
4542 this function has completed hwasan_tagged_stack_vars is empty and all
4543 objects it had pointed to are deallocated. */
4544void
4545hwasan_emit_prologue ()
4546{
4547 /* We need untagged base pointers since libhwasan only accepts untagged
4548 pointers in __hwasan_tag_memory. We need the tagged base pointer to obtain
4549 the base tag for an offset. */
4550
4551 if (hwasan_tagged_stack_vars.is_empty ())
4552 return;
4553
4554 poly_int64 bot = 0, top = 0;
4555 for (hwasan_stack_var &cur : hwasan_tagged_stack_vars)
4556 {
4557 poly_int64 nearest = cur.nearest_offset;
4558 poly_int64 farthest = cur.farthest_offset;
4559
4560 if (known_ge (nearest, farthest))
4561 {
4562 top = nearest;
4563 bot = farthest;
4564 }
4565 else
4566 {
4567 /* Given how these values are calculated, one must be known greater
4568 than the other. */
4569 gcc_assert (known_le (nearest, farthest));
4570 top = farthest;
4571 bot = nearest;
4572 }
4573 poly_int64 size = (top - bot);
4574
4575 /* Assert the edge of each variable is aligned to the HWASAN tag granule
4576 size. */
4577 gcc_assert (multiple_p (top, HWASAN_TAG_GRANULE_SIZE));
4578 gcc_assert (multiple_p (bot, HWASAN_TAG_GRANULE_SIZE));
4579 gcc_assert (multiple_p (size, HWASAN_TAG_GRANULE_SIZE));
4580
4581 rtx fn = init_one_libfunc ("__hwasan_tag_memory");
4582 rtx base_tag = targetm.memtag.extract_tag (cur.tagged_base, NULL_RTX);
4583 rtx tag = plus_constant (QImode, base_tag, cur.tag_offset);
4584 tag = hwasan_truncate_to_tag_size (tag, NULL_RTX);
4585
4586 rtx bottom = convert_memory_address (ptr_mode,
4587 plus_constant (Pmode,
4588 cur.untagged_base,
4589 bot));
4590 emit_library_call (fun: fn, fn_type: LCT_NORMAL, VOIDmode,
4591 arg1: bottom, arg1_mode: ptr_mode,
4592 arg2: tag, QImode,
4593 arg3: gen_int_mode (size, ptr_mode), arg3_mode: ptr_mode);
4594 }
4595 /* Clear the stack vars, we've emitted the prologue for them all now. */
4596 hwasan_tagged_stack_vars.truncate (size: 0);
4597}
4598
4599/* For stack tagging:
4600
4601 Return RTL insns to clear the tags between DYNAMIC and VARS pointers
4602 into the stack. These instructions should be emitted at the end of
4603 every function.
4604
4605 If `dynamic` is NULL_RTX then no insns are returned. */
4606rtx_insn *
4607hwasan_emit_untag_frame (rtx dynamic, rtx vars)
4608{
4609 if (! dynamic)
4610 return NULL;
4611
4612 start_sequence ();
4613
4614 dynamic = convert_memory_address (ptr_mode, dynamic);
4615 vars = convert_memory_address (ptr_mode, vars);
4616
4617 rtx top_rtx;
4618 rtx bot_rtx;
4619 if (FRAME_GROWS_DOWNWARD)
4620 {
4621 top_rtx = vars;
4622 bot_rtx = dynamic;
4623 }
4624 else
4625 {
4626 top_rtx = dynamic;
4627 bot_rtx = vars;
4628 }
4629
4630 rtx size_rtx = expand_simple_binop (ptr_mode, MINUS, top_rtx, bot_rtx,
4631 NULL_RTX, /* unsignedp = */0,
4632 OPTAB_DIRECT);
4633
4634 rtx fn = init_one_libfunc ("__hwasan_tag_memory");
4635 emit_library_call (fun: fn, fn_type: LCT_NORMAL, VOIDmode,
4636 arg1: bot_rtx, arg1_mode: ptr_mode,
4637 HWASAN_STACK_BACKGROUND, QImode,
4638 arg3: size_rtx, arg3_mode: ptr_mode);
4639
4640 do_pending_stack_adjust ();
4641 rtx_insn *insns = get_insns ();
4642 end_sequence ();
4643 return insns;
4644}
4645
4646/* Needs to be GTY(()), because cgraph_build_static_cdtor may
4647 invoke ggc_collect. */
4648static GTY(()) tree hwasan_ctor_statements;
4649
4650/* Insert module initialization into this TU. This initialization calls the
4651 initialization code for libhwasan. */
4652void
4653hwasan_finish_file (void)
4654{
4655 /* Do not emit constructor initialization for the kernel.
4656 (the kernel has its own initialization already). */
4657 if (flag_sanitize & SANITIZE_KERNEL_HWADDRESS)
4658 return;
4659
4660 /* Avoid instrumenting code in the hwasan constructors/destructors. */
4661 flag_sanitize &= ~SANITIZE_HWADDRESS;
4662 int priority = MAX_RESERVED_INIT_PRIORITY - 1;
4663 tree fn = builtin_decl_implicit (fncode: BUILT_IN_HWASAN_INIT);
4664 append_to_statement_list (build_call_expr (fn, 0), &hwasan_ctor_statements);
4665 cgraph_build_static_cdtor (which: 'I', body: hwasan_ctor_statements, priority);
4666 flag_sanitize |= SANITIZE_HWADDRESS;
4667}
4668
4669/* For stack tagging:
4670
4671 Truncate `tag` to the number of bits that a tag uses (i.e. to
4672 HWASAN_TAG_SIZE). Store the result in `target` if it's convenient. */
4673rtx
4674hwasan_truncate_to_tag_size (rtx tag, rtx target)
4675{
4676 gcc_assert (GET_MODE (tag) == QImode);
4677 if (HWASAN_TAG_SIZE != GET_MODE_PRECISION (QImode))
4678 {
4679 gcc_assert (GET_MODE_PRECISION (QImode) > HWASAN_TAG_SIZE);
4680 rtx mask = gen_int_mode ((HOST_WIDE_INT_1U << HWASAN_TAG_SIZE) - 1,
4681 QImode);
4682 tag = expand_simple_binop (QImode, AND, tag, mask, target,
4683 /* unsignedp = */1, OPTAB_WIDEN);
4684 gcc_assert (tag);
4685 }
4686 return tag;
4687}
4688
4689/* Construct a function tree for __hwasan_{load,store}{1,2,4,8,16,_n}.
4690 IS_STORE is either 1 (for a store) or 0 (for a load). */
4691static combined_fn
4692hwasan_check_func (bool is_store, bool recover_p, HOST_WIDE_INT size_in_bytes,
4693 int *nargs)
4694{
4695 static enum built_in_function check[2][2][6]
4696 = { { { BUILT_IN_HWASAN_LOAD1, BUILT_IN_HWASAN_LOAD2,
4697 BUILT_IN_HWASAN_LOAD4, BUILT_IN_HWASAN_LOAD8,
4698 BUILT_IN_HWASAN_LOAD16, BUILT_IN_HWASAN_LOADN },
4699 { BUILT_IN_HWASAN_STORE1, BUILT_IN_HWASAN_STORE2,
4700 BUILT_IN_HWASAN_STORE4, BUILT_IN_HWASAN_STORE8,
4701 BUILT_IN_HWASAN_STORE16, BUILT_IN_HWASAN_STOREN } },
4702 { { BUILT_IN_HWASAN_LOAD1_NOABORT,
4703 BUILT_IN_HWASAN_LOAD2_NOABORT,
4704 BUILT_IN_HWASAN_LOAD4_NOABORT,
4705 BUILT_IN_HWASAN_LOAD8_NOABORT,
4706 BUILT_IN_HWASAN_LOAD16_NOABORT,
4707 BUILT_IN_HWASAN_LOADN_NOABORT },
4708 { BUILT_IN_HWASAN_STORE1_NOABORT,
4709 BUILT_IN_HWASAN_STORE2_NOABORT,
4710 BUILT_IN_HWASAN_STORE4_NOABORT,
4711 BUILT_IN_HWASAN_STORE8_NOABORT,
4712 BUILT_IN_HWASAN_STORE16_NOABORT,
4713 BUILT_IN_HWASAN_STOREN_NOABORT } } };
4714 if (size_in_bytes == -1)
4715 {
4716 *nargs = 2;
4717 return as_combined_fn (fn: check[recover_p][is_store][5]);
4718 }
4719 *nargs = 1;
4720 int size_log2 = exact_log2 (x: size_in_bytes);
4721 gcc_assert (size_log2 >= 0 && size_log2 <= 5);
4722 return as_combined_fn (fn: check[recover_p][is_store][size_log2]);
4723}
4724
4725/* Expand the HWASAN_{LOAD,STORE} builtins. */
4726bool
4727hwasan_expand_check_ifn (gimple_stmt_iterator *iter, bool)
4728{
4729 gimple *g = gsi_stmt (i: *iter);
4730 location_t loc = gimple_location (g);
4731 bool recover_p;
4732 if (flag_sanitize & SANITIZE_USER_HWADDRESS)
4733 recover_p = (flag_sanitize_recover & SANITIZE_USER_HWADDRESS) != 0;
4734 else
4735 recover_p = (flag_sanitize_recover & SANITIZE_KERNEL_HWADDRESS) != 0;
4736
4737 HOST_WIDE_INT flags = tree_to_shwi (gimple_call_arg (gs: g, index: 0));
4738 gcc_assert (flags < ASAN_CHECK_LAST);
4739 bool is_scalar_access = (flags & ASAN_CHECK_SCALAR_ACCESS) != 0;
4740 bool is_store = (flags & ASAN_CHECK_STORE) != 0;
4741 bool is_non_zero_len = (flags & ASAN_CHECK_NON_ZERO_LEN) != 0;
4742
4743 tree base = gimple_call_arg (gs: g, index: 1);
4744 tree len = gimple_call_arg (gs: g, index: 2);
4745
4746 /* `align` is unused for HWASAN_CHECK, but we pass the argument anyway
4747 since that way the arguments match ASAN_CHECK. */
4748 /* HOST_WIDE_INT align = tree_to_shwi (gimple_call_arg (g, 3)); */
4749
4750 unsigned HOST_WIDE_INT size_in_bytes
4751 = is_scalar_access ? tree_to_shwi (len) : -1;
4752
4753 gimple_stmt_iterator gsi = *iter;
4754
4755 if (!is_non_zero_len)
4756 {
4757 /* So, the length of the memory area to hwasan-protect is
4758 non-constant. Let's guard the generated instrumentation code
4759 like:
4760
4761 if (len != 0)
4762 {
4763 // hwasan instrumentation code goes here.
4764 }
4765 // falltrough instructions, starting with *ITER. */
4766
4767 g = gimple_build_cond (NE_EXPR,
4768 len,
4769 build_int_cst (TREE_TYPE (len), 0),
4770 NULL_TREE, NULL_TREE);
4771 gimple_set_location (g, location: loc);
4772
4773 basic_block then_bb, fallthrough_bb;
4774 insert_if_then_before_iter (cond: as_a <gcond *> (p: g), iter,
4775 /*then_more_likely_p=*/true,
4776 then_bb: &then_bb, fallthrough_bb: &fallthrough_bb);
4777 /* Note that fallthrough_bb starts with the statement that was
4778 pointed to by ITER. */
4779
4780 /* The 'then block' of the 'if (len != 0) condition is where
4781 we'll generate the hwasan instrumentation code now. */
4782 gsi = gsi_last_bb (bb: then_bb);
4783 }
4784
4785 gimple_seq stmts = NULL;
4786 tree base_addr = gimple_build (seq: &stmts, loc, code: NOP_EXPR,
4787 pointer_sized_int_node, ops: base);
4788
4789 int nargs = 0;
4790 combined_fn fn
4791 = hwasan_check_func (is_store, recover_p, size_in_bytes, nargs: &nargs);
4792 if (nargs == 1)
4793 gimple_build (seq: &stmts, loc, fn, void_type_node, args: base_addr);
4794 else
4795 {
4796 gcc_assert (nargs == 2);
4797 tree sz_arg = gimple_build (seq: &stmts, loc, code: NOP_EXPR,
4798 pointer_sized_int_node, ops: len);
4799 gimple_build (seq: &stmts, loc, fn, void_type_node, args: base_addr, args: sz_arg);
4800 }
4801
4802 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
4803 gsi_remove (iter, true);
4804 *iter = gsi;
4805 return false;
4806}
4807
4808/* For stack tagging:
4809
4810 Dummy: the HWASAN_MARK internal function should only ever be in the code
4811 after the sanopt pass. */
4812bool
4813hwasan_expand_mark_ifn (gimple_stmt_iterator *)
4814{
4815 gcc_unreachable ();
4816}
4817
4818bool
4819gate_hwasan ()
4820{
4821 return hwasan_sanitize_p ();
4822}
4823
4824#include "gt-asan.h"
4825

source code of gcc/asan.cc