1 //===- AddressSanitizer.cpp - memory error detector -----------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file is a part of AddressSanitizer, an address sanity checker.
10 // Details of the algorithm:
11 // https://github.com/google/sanitizers/wiki/AddressSanitizerAlgorithm
12 //
13 // FIXME: This sanitizer does not yet handle scalable vectors
14 //
15 //===----------------------------------------------------------------------===//
16
17 #include "llvm/Transforms/Instrumentation/AddressSanitizer.h"
18 #include "llvm/ADT/ArrayRef.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/ADT/StringRef.h"
26 #include "llvm/ADT/Triple.h"
27 #include "llvm/ADT/Twine.h"
28 #include "llvm/Analysis/MemoryBuiltins.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/BinaryFormat/MachO.h"
32 #include "llvm/IR/Argument.h"
33 #include "llvm/IR/Attributes.h"
34 #include "llvm/IR/BasicBlock.h"
35 #include "llvm/IR/Comdat.h"
36 #include "llvm/IR/Constant.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DIBuilder.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DebugInfoMetadata.h"
41 #include "llvm/IR/DebugLoc.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/Dominators.h"
44 #include "llvm/IR/Function.h"
45 #include "llvm/IR/GlobalAlias.h"
46 #include "llvm/IR/GlobalValue.h"
47 #include "llvm/IR/GlobalVariable.h"
48 #include "llvm/IR/IRBuilder.h"
49 #include "llvm/IR/InlineAsm.h"
50 #include "llvm/IR/InstVisitor.h"
51 #include "llvm/IR/InstrTypes.h"
52 #include "llvm/IR/Instruction.h"
53 #include "llvm/IR/Instructions.h"
54 #include "llvm/IR/IntrinsicInst.h"
55 #include "llvm/IR/Intrinsics.h"
56 #include "llvm/IR/LLVMContext.h"
57 #include "llvm/IR/MDBuilder.h"
58 #include "llvm/IR/Metadata.h"
59 #include "llvm/IR/Module.h"
60 #include "llvm/IR/Type.h"
61 #include "llvm/IR/Use.h"
62 #include "llvm/IR/Value.h"
63 #include "llvm/InitializePasses.h"
64 #include "llvm/MC/MCSectionMachO.h"
65 #include "llvm/Pass.h"
66 #include "llvm/Support/Casting.h"
67 #include "llvm/Support/CommandLine.h"
68 #include "llvm/Support/Debug.h"
69 #include "llvm/Support/ErrorHandling.h"
70 #include "llvm/Support/MathExtras.h"
71 #include "llvm/Support/ScopedPrinter.h"
72 #include "llvm/Support/raw_ostream.h"
73 #include "llvm/Transforms/Instrumentation.h"
74 #include "llvm/Transforms/Instrumentation/AddressSanitizerCommon.h"
75 #include "llvm/Transforms/Utils/ASanStackFrameLayout.h"
76 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
77 #include "llvm/Transforms/Utils/Local.h"
78 #include "llvm/Transforms/Utils/ModuleUtils.h"
79 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
80 #include <algorithm>
81 #include <cassert>
82 #include <cstddef>
83 #include <cstdint>
84 #include <iomanip>
85 #include <limits>
86 #include <memory>
87 #include <sstream>
88 #include <string>
89 #include <tuple>
90
91 using namespace llvm;
92
93 #define DEBUG_TYPE "asan"
94
95 static const uint64_t kDefaultShadowScale = 3;
96 static const uint64_t kDefaultShadowOffset32 = 1ULL << 29;
97 static const uint64_t kDefaultShadowOffset64 = 1ULL << 44;
98 static const uint64_t kDynamicShadowSentinel =
99 std::numeric_limits<uint64_t>::max();
100 static const uint64_t kSmallX86_64ShadowOffsetBase = 0x7FFFFFFF; // < 2G.
101 static const uint64_t kSmallX86_64ShadowOffsetAlignMask = ~0xFFFULL;
102 static const uint64_t kLinuxKasan_ShadowOffset64 = 0xdffffc0000000000;
103 static const uint64_t kPPC64_ShadowOffset64 = 1ULL << 44;
104 static const uint64_t kSystemZ_ShadowOffset64 = 1ULL << 52;
105 static const uint64_t kMIPS32_ShadowOffset32 = 0x0aaa0000;
106 static const uint64_t kMIPS64_ShadowOffset64 = 1ULL << 37;
107 static const uint64_t kAArch64_ShadowOffset64 = 1ULL << 36;
108 static const uint64_t kRISCV64_ShadowOffset64 = 0x20000000;
109 static const uint64_t kFreeBSD_ShadowOffset32 = 1ULL << 30;
110 static const uint64_t kFreeBSD_ShadowOffset64 = 1ULL << 46;
111 static const uint64_t kFreeBSDKasan_ShadowOffset64 = 0xdffff7c000000000;
112 static const uint64_t kNetBSD_ShadowOffset32 = 1ULL << 30;
113 static const uint64_t kNetBSD_ShadowOffset64 = 1ULL << 46;
114 static const uint64_t kNetBSDKasan_ShadowOffset64 = 0xdfff900000000000;
115 static const uint64_t kPS4CPU_ShadowOffset64 = 1ULL << 40;
116 static const uint64_t kWindowsShadowOffset32 = 3ULL << 28;
117 static const uint64_t kEmscriptenShadowOffset = 0;
118
119 static const uint64_t kMyriadShadowScale = 5;
120 static const uint64_t kMyriadMemoryOffset32 = 0x80000000ULL;
121 static const uint64_t kMyriadMemorySize32 = 0x20000000ULL;
122 static const uint64_t kMyriadTagShift = 29;
123 static const uint64_t kMyriadDDRTag = 4;
124 static const uint64_t kMyriadCacheBitMask32 = 0x40000000ULL;
125
126 // The shadow memory space is dynamically allocated.
127 static const uint64_t kWindowsShadowOffset64 = kDynamicShadowSentinel;
128
129 static const size_t kMinStackMallocSize = 1 << 6; // 64B
130 static const size_t kMaxStackMallocSize = 1 << 16; // 64K
131 static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3;
132 static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E;
133
134 const char kAsanModuleCtorName[] = "asan.module_ctor";
135 const char kAsanModuleDtorName[] = "asan.module_dtor";
136 static const uint64_t kAsanCtorAndDtorPriority = 1;
137 // On Emscripten, the system needs more than one priorities for constructors.
138 static const uint64_t kAsanEmscriptenCtorAndDtorPriority = 50;
139 const char kAsanReportErrorTemplate[] = "__asan_report_";
140 const char kAsanRegisterGlobalsName[] = "__asan_register_globals";
141 const char kAsanUnregisterGlobalsName[] = "__asan_unregister_globals";
142 const char kAsanRegisterImageGlobalsName[] = "__asan_register_image_globals";
143 const char kAsanUnregisterImageGlobalsName[] =
144 "__asan_unregister_image_globals";
145 const char kAsanRegisterElfGlobalsName[] = "__asan_register_elf_globals";
146 const char kAsanUnregisterElfGlobalsName[] = "__asan_unregister_elf_globals";
147 const char kAsanPoisonGlobalsName[] = "__asan_before_dynamic_init";
148 const char kAsanUnpoisonGlobalsName[] = "__asan_after_dynamic_init";
149 const char kAsanInitName[] = "__asan_init";
150 const char kAsanVersionCheckNamePrefix[] = "__asan_version_mismatch_check_v";
151 const char kAsanPtrCmp[] = "__sanitizer_ptr_cmp";
152 const char kAsanPtrSub[] = "__sanitizer_ptr_sub";
153 const char kAsanHandleNoReturnName[] = "__asan_handle_no_return";
154 static const int kMaxAsanStackMallocSizeClass = 10;
155 const char kAsanStackMallocNameTemplate[] = "__asan_stack_malloc_";
156 const char kAsanStackFreeNameTemplate[] = "__asan_stack_free_";
157 const char kAsanGenPrefix[] = "___asan_gen_";
158 const char kODRGenPrefix[] = "__odr_asan_gen_";
159 const char kSanCovGenPrefix[] = "__sancov_gen_";
160 const char kAsanSetShadowPrefix[] = "__asan_set_shadow_";
161 const char kAsanPoisonStackMemoryName[] = "__asan_poison_stack_memory";
162 const char kAsanUnpoisonStackMemoryName[] = "__asan_unpoison_stack_memory";
163
164 // ASan version script has __asan_* wildcard. Triple underscore prevents a
165 // linker (gold) warning about attempting to export a local symbol.
166 const char kAsanGlobalsRegisteredFlagName[] = "___asan_globals_registered";
167
168 const char kAsanOptionDetectUseAfterReturn[] =
169 "__asan_option_detect_stack_use_after_return";
170
171 const char kAsanShadowMemoryDynamicAddress[] =
172 "__asan_shadow_memory_dynamic_address";
173
174 const char kAsanAllocaPoison[] = "__asan_alloca_poison";
175 const char kAsanAllocasUnpoison[] = "__asan_allocas_unpoison";
176
177 // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
178 static const size_t kNumberOfAccessSizes = 5;
179
180 static const unsigned kAllocaRzSize = 32;
181
182 // Command-line flags.
183
184 static cl::opt<bool> ClEnableKasan(
185 "asan-kernel", cl::desc("Enable KernelAddressSanitizer instrumentation"),
186 cl::Hidden, cl::init(false));
187
188 static cl::opt<bool> ClRecover(
189 "asan-recover",
190 cl::desc("Enable recovery mode (continue-after-error)."),
191 cl::Hidden, cl::init(false));
192
193 static cl::opt<bool> ClInsertVersionCheck(
194 "asan-guard-against-version-mismatch",
195 cl::desc("Guard against compiler/runtime version mismatch."),
196 cl::Hidden, cl::init(true));
197
198 // This flag may need to be replaced with -f[no-]asan-reads.
199 static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
200 cl::desc("instrument read instructions"),
201 cl::Hidden, cl::init(true));
202
203 static cl::opt<bool> ClInstrumentWrites(
204 "asan-instrument-writes", cl::desc("instrument write instructions"),
205 cl::Hidden, cl::init(true));
206
207 static cl::opt<bool> ClInstrumentAtomics(
208 "asan-instrument-atomics",
209 cl::desc("instrument atomic instructions (rmw, cmpxchg)"), cl::Hidden,
210 cl::init(true));
211
212 static cl::opt<bool>
213 ClInstrumentByval("asan-instrument-byval",
214 cl::desc("instrument byval call arguments"), cl::Hidden,
215 cl::init(true));
216
217 static cl::opt<bool> ClAlwaysSlowPath(
218 "asan-always-slow-path",
219 cl::desc("use instrumentation with slow path for all accesses"), cl::Hidden,
220 cl::init(false));
221
222 static cl::opt<bool> ClForceDynamicShadow(
223 "asan-force-dynamic-shadow",
224 cl::desc("Load shadow address into a local variable for each function"),
225 cl::Hidden, cl::init(false));
226
227 static cl::opt<bool>
228 ClWithIfunc("asan-with-ifunc",
229 cl::desc("Access dynamic shadow through an ifunc global on "
230 "platforms that support this"),
231 cl::Hidden, cl::init(true));
232
233 static cl::opt<bool> ClWithIfuncSuppressRemat(
234 "asan-with-ifunc-suppress-remat",
235 cl::desc("Suppress rematerialization of dynamic shadow address by passing "
236 "it through inline asm in prologue."),
237 cl::Hidden, cl::init(true));
238
239 // This flag limits the number of instructions to be instrumented
240 // in any given BB. Normally, this should be set to unlimited (INT_MAX),
241 // but due to http://llvm.org/bugs/show_bug.cgi?id=12652 we temporary
242 // set it to 10000.
243 static cl::opt<int> ClMaxInsnsToInstrumentPerBB(
244 "asan-max-ins-per-bb", cl::init(10000),
245 cl::desc("maximal number of instructions to instrument in any given BB"),
246 cl::Hidden);
247
248 // This flag may need to be replaced with -f[no]asan-stack.
249 static cl::opt<bool> ClStack("asan-stack", cl::desc("Handle stack memory"),
250 cl::Hidden, cl::init(true));
251 static cl::opt<uint32_t> ClMaxInlinePoisoningSize(
252 "asan-max-inline-poisoning-size",
253 cl::desc(
254 "Inline shadow poisoning for blocks up to the given size in bytes."),
255 cl::Hidden, cl::init(64));
256
257 static cl::opt<bool> ClUseAfterReturn("asan-use-after-return",
258 cl::desc("Check stack-use-after-return"),
259 cl::Hidden, cl::init(true));
260
261 static cl::opt<bool> ClRedzoneByvalArgs("asan-redzone-byval-args",
262 cl::desc("Create redzones for byval "
263 "arguments (extra copy "
264 "required)"), cl::Hidden,
265 cl::init(true));
266
267 static cl::opt<bool> ClUseAfterScope("asan-use-after-scope",
268 cl::desc("Check stack-use-after-scope"),
269 cl::Hidden, cl::init(false));
270
271 // This flag may need to be replaced with -f[no]asan-globals.
272 static cl::opt<bool> ClGlobals("asan-globals",
273 cl::desc("Handle global objects"), cl::Hidden,
274 cl::init(true));
275
276 static cl::opt<bool> ClInitializers("asan-initialization-order",
277 cl::desc("Handle C++ initializer order"),
278 cl::Hidden, cl::init(true));
279
280 static cl::opt<bool> ClInvalidPointerPairs(
281 "asan-detect-invalid-pointer-pair",
282 cl::desc("Instrument <, <=, >, >=, - with pointer operands"), cl::Hidden,
283 cl::init(false));
284
285 static cl::opt<bool> ClInvalidPointerCmp(
286 "asan-detect-invalid-pointer-cmp",
287 cl::desc("Instrument <, <=, >, >= with pointer operands"), cl::Hidden,
288 cl::init(false));
289
290 static cl::opt<bool> ClInvalidPointerSub(
291 "asan-detect-invalid-pointer-sub",
292 cl::desc("Instrument - operations with pointer operands"), cl::Hidden,
293 cl::init(false));
294
295 static cl::opt<unsigned> ClRealignStack(
296 "asan-realign-stack",
297 cl::desc("Realign stack to the value of this flag (power of two)"),
298 cl::Hidden, cl::init(32));
299
300 static cl::opt<int> ClInstrumentationWithCallsThreshold(
301 "asan-instrumentation-with-call-threshold",
302 cl::desc(
303 "If the function being instrumented contains more than "
304 "this number of memory accesses, use callbacks instead of "
305 "inline checks (-1 means never use callbacks)."),
306 cl::Hidden, cl::init(7000));
307
308 static cl::opt<std::string> ClMemoryAccessCallbackPrefix(
309 "asan-memory-access-callback-prefix",
310 cl::desc("Prefix for memory access callbacks"), cl::Hidden,
311 cl::init("__asan_"));
312
313 static cl::opt<bool>
314 ClInstrumentDynamicAllocas("asan-instrument-dynamic-allocas",
315 cl::desc("instrument dynamic allocas"),
316 cl::Hidden, cl::init(true));
317
318 static cl::opt<bool> ClSkipPromotableAllocas(
319 "asan-skip-promotable-allocas",
320 cl::desc("Do not instrument promotable allocas"), cl::Hidden,
321 cl::init(true));
322
323 // These flags allow to change the shadow mapping.
324 // The shadow mapping looks like
325 // Shadow = (Mem >> scale) + offset
326
327 static cl::opt<int> ClMappingScale("asan-mapping-scale",
328 cl::desc("scale of asan shadow mapping"),
329 cl::Hidden, cl::init(0));
330
331 static cl::opt<uint64_t>
332 ClMappingOffset("asan-mapping-offset",
333 cl::desc("offset of asan shadow mapping [EXPERIMENTAL]"),
334 cl::Hidden, cl::init(0));
335
336 // Optimization flags. Not user visible, used mostly for testing
337 // and benchmarking the tool.
338
339 static cl::opt<bool> ClOpt("asan-opt", cl::desc("Optimize instrumentation"),
340 cl::Hidden, cl::init(true));
341
342 static cl::opt<bool> ClOptSameTemp(
343 "asan-opt-same-temp", cl::desc("Instrument the same temp just once"),
344 cl::Hidden, cl::init(true));
345
346 static cl::opt<bool> ClOptGlobals("asan-opt-globals",
347 cl::desc("Don't instrument scalar globals"),
348 cl::Hidden, cl::init(true));
349
350 static cl::opt<bool> ClOptStack(
351 "asan-opt-stack", cl::desc("Don't instrument scalar stack variables"),
352 cl::Hidden, cl::init(false));
353
354 static cl::opt<bool> ClDynamicAllocaStack(
355 "asan-stack-dynamic-alloca",
356 cl::desc("Use dynamic alloca to represent stack variables"), cl::Hidden,
357 cl::init(true));
358
359 static cl::opt<uint32_t> ClForceExperiment(
360 "asan-force-experiment",
361 cl::desc("Force optimization experiment (for testing)"), cl::Hidden,
362 cl::init(0));
363
364 static cl::opt<bool>
365 ClUsePrivateAlias("asan-use-private-alias",
366 cl::desc("Use private aliases for global variables"),
367 cl::Hidden, cl::init(false));
368
369 static cl::opt<bool>
370 ClUseOdrIndicator("asan-use-odr-indicator",
371 cl::desc("Use odr indicators to improve ODR reporting"),
372 cl::Hidden, cl::init(false));
373
374 static cl::opt<bool>
375 ClUseGlobalsGC("asan-globals-live-support",
376 cl::desc("Use linker features to support dead "
377 "code stripping of globals"),
378 cl::Hidden, cl::init(true));
379
380 // This is on by default even though there is a bug in gold:
381 // https://sourceware.org/bugzilla/show_bug.cgi?id=19002
382 static cl::opt<bool>
383 ClWithComdat("asan-with-comdat",
384 cl::desc("Place ASan constructors in comdat sections"),
385 cl::Hidden, cl::init(true));
386
387 // Debug flags.
388
389 static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
390 cl::init(0));
391
392 static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
393 cl::Hidden, cl::init(0));
394
395 static cl::opt<std::string> ClDebugFunc("asan-debug-func", cl::Hidden,
396 cl::desc("Debug func"));
397
398 static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
399 cl::Hidden, cl::init(-1));
400
401 static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug max inst"),
402 cl::Hidden, cl::init(-1));
403
404 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
405 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
406 STATISTIC(NumOptimizedAccessesToGlobalVar,
407 "Number of optimized accesses to global vars");
408 STATISTIC(NumOptimizedAccessesToStackVar,
409 "Number of optimized accesses to stack vars");
410
411 namespace {
412
413 /// This struct defines the shadow mapping using the rule:
414 /// shadow = (mem >> Scale) ADD-or-OR Offset.
415 /// If InGlobal is true, then
416 /// extern char __asan_shadow[];
417 /// shadow = (mem >> Scale) + &__asan_shadow
418 struct ShadowMapping {
419 int Scale;
420 uint64_t Offset;
421 bool OrShadowOffset;
422 bool InGlobal;
423 };
424
425 } // end anonymous namespace
426
getShadowMapping(Triple & TargetTriple,int LongSize,bool IsKasan)427 static ShadowMapping getShadowMapping(Triple &TargetTriple, int LongSize,
428 bool IsKasan) {
429 bool IsAndroid = TargetTriple.isAndroid();
430 bool IsIOS = TargetTriple.isiOS() || TargetTriple.isWatchOS();
431 bool IsMacOS = TargetTriple.isMacOSX();
432 bool IsFreeBSD = TargetTriple.isOSFreeBSD();
433 bool IsNetBSD = TargetTriple.isOSNetBSD();
434 bool IsPS4CPU = TargetTriple.isPS4CPU();
435 bool IsLinux = TargetTriple.isOSLinux();
436 bool IsPPC64 = TargetTriple.getArch() == Triple::ppc64 ||
437 TargetTriple.getArch() == Triple::ppc64le;
438 bool IsSystemZ = TargetTriple.getArch() == Triple::systemz;
439 bool IsX86_64 = TargetTriple.getArch() == Triple::x86_64;
440 bool IsMIPS32 = TargetTriple.isMIPS32();
441 bool IsMIPS64 = TargetTriple.isMIPS64();
442 bool IsArmOrThumb = TargetTriple.isARM() || TargetTriple.isThumb();
443 bool IsAArch64 = TargetTriple.getArch() == Triple::aarch64;
444 bool IsRISCV64 = TargetTriple.getArch() == Triple::riscv64;
445 bool IsWindows = TargetTriple.isOSWindows();
446 bool IsFuchsia = TargetTriple.isOSFuchsia();
447 bool IsMyriad = TargetTriple.getVendor() == llvm::Triple::Myriad;
448 bool IsEmscripten = TargetTriple.isOSEmscripten();
449
450 ShadowMapping Mapping;
451
452 Mapping.Scale = IsMyriad ? kMyriadShadowScale : kDefaultShadowScale;
453 if (ClMappingScale.getNumOccurrences() > 0) {
454 Mapping.Scale = ClMappingScale;
455 }
456
457 if (LongSize == 32) {
458 if (IsAndroid)
459 Mapping.Offset = kDynamicShadowSentinel;
460 else if (IsMIPS32)
461 Mapping.Offset = kMIPS32_ShadowOffset32;
462 else if (IsFreeBSD)
463 Mapping.Offset = kFreeBSD_ShadowOffset32;
464 else if (IsNetBSD)
465 Mapping.Offset = kNetBSD_ShadowOffset32;
466 else if (IsIOS)
467 Mapping.Offset = kDynamicShadowSentinel;
468 else if (IsWindows)
469 Mapping.Offset = kWindowsShadowOffset32;
470 else if (IsEmscripten)
471 Mapping.Offset = kEmscriptenShadowOffset;
472 else if (IsMyriad) {
473 uint64_t ShadowOffset = (kMyriadMemoryOffset32 + kMyriadMemorySize32 -
474 (kMyriadMemorySize32 >> Mapping.Scale));
475 Mapping.Offset = ShadowOffset - (kMyriadMemoryOffset32 >> Mapping.Scale);
476 }
477 else
478 Mapping.Offset = kDefaultShadowOffset32;
479 } else { // LongSize == 64
480 // Fuchsia is always PIE, which means that the beginning of the address
481 // space is always available.
482 if (IsFuchsia)
483 Mapping.Offset = 0;
484 else if (IsPPC64)
485 Mapping.Offset = kPPC64_ShadowOffset64;
486 else if (IsSystemZ)
487 Mapping.Offset = kSystemZ_ShadowOffset64;
488 else if (IsFreeBSD && !IsMIPS64) {
489 if (IsKasan)
490 Mapping.Offset = kFreeBSDKasan_ShadowOffset64;
491 else
492 Mapping.Offset = kFreeBSD_ShadowOffset64;
493 } else if (IsNetBSD) {
494 if (IsKasan)
495 Mapping.Offset = kNetBSDKasan_ShadowOffset64;
496 else
497 Mapping.Offset = kNetBSD_ShadowOffset64;
498 } else if (IsPS4CPU)
499 Mapping.Offset = kPS4CPU_ShadowOffset64;
500 else if (IsLinux && IsX86_64) {
501 if (IsKasan)
502 Mapping.Offset = kLinuxKasan_ShadowOffset64;
503 else
504 Mapping.Offset = (kSmallX86_64ShadowOffsetBase &
505 (kSmallX86_64ShadowOffsetAlignMask << Mapping.Scale));
506 } else if (IsWindows && IsX86_64) {
507 Mapping.Offset = kWindowsShadowOffset64;
508 } else if (IsMIPS64)
509 Mapping.Offset = kMIPS64_ShadowOffset64;
510 else if (IsIOS)
511 Mapping.Offset = kDynamicShadowSentinel;
512 else if (IsMacOS && IsAArch64)
513 Mapping.Offset = kDynamicShadowSentinel;
514 else if (IsAArch64)
515 Mapping.Offset = kAArch64_ShadowOffset64;
516 else if (IsRISCV64)
517 Mapping.Offset = kRISCV64_ShadowOffset64;
518 else
519 Mapping.Offset = kDefaultShadowOffset64;
520 }
521
522 if (ClForceDynamicShadow) {
523 Mapping.Offset = kDynamicShadowSentinel;
524 }
525
526 if (ClMappingOffset.getNumOccurrences() > 0) {
527 Mapping.Offset = ClMappingOffset;
528 }
529
530 // OR-ing shadow offset if more efficient (at least on x86) if the offset
531 // is a power of two, but on ppc64 we have to use add since the shadow
532 // offset is not necessary 1/8-th of the address space. On SystemZ,
533 // we could OR the constant in a single instruction, but it's more
534 // efficient to load it once and use indexed addressing.
535 Mapping.OrShadowOffset = !IsAArch64 && !IsPPC64 && !IsSystemZ && !IsPS4CPU &&
536 !IsRISCV64 &&
537 !(Mapping.Offset & (Mapping.Offset - 1)) &&
538 Mapping.Offset != kDynamicShadowSentinel;
539 bool IsAndroidWithIfuncSupport =
540 IsAndroid && !TargetTriple.isAndroidVersionLT(21);
541 Mapping.InGlobal = ClWithIfunc && IsAndroidWithIfuncSupport && IsArmOrThumb;
542
543 return Mapping;
544 }
545
getRedzoneSizeForScale(int MappingScale)546 static uint64_t getRedzoneSizeForScale(int MappingScale) {
547 // Redzone used for stack and globals is at least 32 bytes.
548 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
549 return std::max(32U, 1U << MappingScale);
550 }
551
GetCtorAndDtorPriority(Triple & TargetTriple)552 static uint64_t GetCtorAndDtorPriority(Triple &TargetTriple) {
553 if (TargetTriple.isOSEmscripten()) {
554 return kAsanEmscriptenCtorAndDtorPriority;
555 } else {
556 return kAsanCtorAndDtorPriority;
557 }
558 }
559
560 namespace {
561
562 /// Module analysis for getting various metadata about the module.
563 class ASanGlobalsMetadataWrapperPass : public ModulePass {
564 public:
565 static char ID;
566
ASanGlobalsMetadataWrapperPass()567 ASanGlobalsMetadataWrapperPass() : ModulePass(ID) {
568 initializeASanGlobalsMetadataWrapperPassPass(
569 *PassRegistry::getPassRegistry());
570 }
571
runOnModule(Module & M)572 bool runOnModule(Module &M) override {
573 GlobalsMD = GlobalsMetadata(M);
574 return false;
575 }
576
getPassName() const577 StringRef getPassName() const override {
578 return "ASanGlobalsMetadataWrapperPass";
579 }
580
getAnalysisUsage(AnalysisUsage & AU) const581 void getAnalysisUsage(AnalysisUsage &AU) const override {
582 AU.setPreservesAll();
583 }
584
getGlobalsMD()585 GlobalsMetadata &getGlobalsMD() { return GlobalsMD; }
586
587 private:
588 GlobalsMetadata GlobalsMD;
589 };
590
591 char ASanGlobalsMetadataWrapperPass::ID = 0;
592
593 /// AddressSanitizer: instrument the code in module to find memory bugs.
594 struct AddressSanitizer {
AddressSanitizer__anonfe46d5c80211::AddressSanitizer595 AddressSanitizer(Module &M, const GlobalsMetadata *GlobalsMD,
596 bool CompileKernel = false, bool Recover = false,
597 bool UseAfterScope = false)
598 : CompileKernel(ClEnableKasan.getNumOccurrences() > 0 ? ClEnableKasan
599 : CompileKernel),
600 Recover(ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover),
601 UseAfterScope(UseAfterScope || ClUseAfterScope), GlobalsMD(*GlobalsMD) {
602 C = &(M.getContext());
603 LongSize = M.getDataLayout().getPointerSizeInBits();
604 IntptrTy = Type::getIntNTy(*C, LongSize);
605 TargetTriple = Triple(M.getTargetTriple());
606
607 Mapping = getShadowMapping(TargetTriple, LongSize, this->CompileKernel);
608 }
609
getAllocaSizeInBytes__anonfe46d5c80211::AddressSanitizer610 uint64_t getAllocaSizeInBytes(const AllocaInst &AI) const {
611 uint64_t ArraySize = 1;
612 if (AI.isArrayAllocation()) {
613 const ConstantInt *CI = dyn_cast<ConstantInt>(AI.getArraySize());
614 assert(CI && "non-constant array size");
615 ArraySize = CI->getZExtValue();
616 }
617 Type *Ty = AI.getAllocatedType();
618 uint64_t SizeInBytes =
619 AI.getModule()->getDataLayout().getTypeAllocSize(Ty);
620 return SizeInBytes * ArraySize;
621 }
622
623 /// Check if we want (and can) handle this alloca.
624 bool isInterestingAlloca(const AllocaInst &AI);
625
626 bool ignoreAccess(Value *Ptr);
627 void getInterestingMemoryOperands(
628 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting);
629
630 void instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
631 InterestingMemoryOperand &O, bool UseCalls,
632 const DataLayout &DL);
633 void instrumentPointerComparisonOrSubtraction(Instruction *I);
634 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore,
635 Value *Addr, uint32_t TypeSize, bool IsWrite,
636 Value *SizeArgument, bool UseCalls, uint32_t Exp);
637 void instrumentUnusualSizeOrAlignment(Instruction *I,
638 Instruction *InsertBefore, Value *Addr,
639 uint32_t TypeSize, bool IsWrite,
640 Value *SizeArgument, bool UseCalls,
641 uint32_t Exp);
642 Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
643 Value *ShadowValue, uint32_t TypeSize);
644 Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr,
645 bool IsWrite, size_t AccessSizeIndex,
646 Value *SizeArgument, uint32_t Exp);
647 void instrumentMemIntrinsic(MemIntrinsic *MI);
648 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
649 bool suppressInstrumentationSiteForDebug(int &Instrumented);
650 bool instrumentFunction(Function &F, const TargetLibraryInfo *TLI);
651 bool maybeInsertAsanInitAtFunctionEntry(Function &F);
652 bool maybeInsertDynamicShadowAtFunctionEntry(Function &F);
653 void markEscapedLocalAllocas(Function &F);
654
655 private:
656 friend struct FunctionStackPoisoner;
657
658 void initializeCallbacks(Module &M);
659
660 bool LooksLikeCodeInBug11395(Instruction *I);
661 bool GlobalIsLinkerInitialized(GlobalVariable *G);
662 bool isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis, Value *Addr,
663 uint64_t TypeSize) const;
664
665 /// Helper to cleanup per-function state.
666 struct FunctionStateRAII {
667 AddressSanitizer *Pass;
668
FunctionStateRAII__anonfe46d5c80211::AddressSanitizer::FunctionStateRAII669 FunctionStateRAII(AddressSanitizer *Pass) : Pass(Pass) {
670 assert(Pass->ProcessedAllocas.empty() &&
671 "last pass forgot to clear cache");
672 assert(!Pass->LocalDynamicShadow);
673 }
674
~FunctionStateRAII__anonfe46d5c80211::AddressSanitizer::FunctionStateRAII675 ~FunctionStateRAII() {
676 Pass->LocalDynamicShadow = nullptr;
677 Pass->ProcessedAllocas.clear();
678 }
679 };
680
681 LLVMContext *C;
682 Triple TargetTriple;
683 int LongSize;
684 bool CompileKernel;
685 bool Recover;
686 bool UseAfterScope;
687 Type *IntptrTy;
688 ShadowMapping Mapping;
689 FunctionCallee AsanHandleNoReturnFunc;
690 FunctionCallee AsanPtrCmpFunction, AsanPtrSubFunction;
691 Constant *AsanShadowGlobal;
692
693 // These arrays is indexed by AccessIsWrite, Experiment and log2(AccessSize).
694 FunctionCallee AsanErrorCallback[2][2][kNumberOfAccessSizes];
695 FunctionCallee AsanMemoryAccessCallback[2][2][kNumberOfAccessSizes];
696
697 // These arrays is indexed by AccessIsWrite and Experiment.
698 FunctionCallee AsanErrorCallbackSized[2][2];
699 FunctionCallee AsanMemoryAccessCallbackSized[2][2];
700
701 FunctionCallee AsanMemmove, AsanMemcpy, AsanMemset;
702 Value *LocalDynamicShadow = nullptr;
703 const GlobalsMetadata &GlobalsMD;
704 DenseMap<const AllocaInst *, bool> ProcessedAllocas;
705 };
706
707 class AddressSanitizerLegacyPass : public FunctionPass {
708 public:
709 static char ID;
710
AddressSanitizerLegacyPass(bool CompileKernel=false,bool Recover=false,bool UseAfterScope=false)711 explicit AddressSanitizerLegacyPass(bool CompileKernel = false,
712 bool Recover = false,
713 bool UseAfterScope = false)
714 : FunctionPass(ID), CompileKernel(CompileKernel), Recover(Recover),
715 UseAfterScope(UseAfterScope) {
716 initializeAddressSanitizerLegacyPassPass(*PassRegistry::getPassRegistry());
717 }
718
getPassName() const719 StringRef getPassName() const override {
720 return "AddressSanitizerFunctionPass";
721 }
722
getAnalysisUsage(AnalysisUsage & AU) const723 void getAnalysisUsage(AnalysisUsage &AU) const override {
724 AU.addRequired<ASanGlobalsMetadataWrapperPass>();
725 AU.addRequired<TargetLibraryInfoWrapperPass>();
726 }
727
runOnFunction(Function & F)728 bool runOnFunction(Function &F) override {
729 GlobalsMetadata &GlobalsMD =
730 getAnalysis<ASanGlobalsMetadataWrapperPass>().getGlobalsMD();
731 const TargetLibraryInfo *TLI =
732 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
733 AddressSanitizer ASan(*F.getParent(), &GlobalsMD, CompileKernel, Recover,
734 UseAfterScope);
735 return ASan.instrumentFunction(F, TLI);
736 }
737
738 private:
739 bool CompileKernel;
740 bool Recover;
741 bool UseAfterScope;
742 };
743
744 class ModuleAddressSanitizer {
745 public:
ModuleAddressSanitizer(Module & M,const GlobalsMetadata * GlobalsMD,bool CompileKernel=false,bool Recover=false,bool UseGlobalsGC=true,bool UseOdrIndicator=false)746 ModuleAddressSanitizer(Module &M, const GlobalsMetadata *GlobalsMD,
747 bool CompileKernel = false, bool Recover = false,
748 bool UseGlobalsGC = true, bool UseOdrIndicator = false)
749 : GlobalsMD(*GlobalsMD),
750 CompileKernel(ClEnableKasan.getNumOccurrences() > 0 ? ClEnableKasan
751 : CompileKernel),
752 Recover(ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover),
753 UseGlobalsGC(UseGlobalsGC && ClUseGlobalsGC && !this->CompileKernel),
754 // Enable aliases as they should have no downside with ODR indicators.
755 UsePrivateAlias(UseOdrIndicator || ClUsePrivateAlias),
756 UseOdrIndicator(UseOdrIndicator || ClUseOdrIndicator),
757 // Not a typo: ClWithComdat is almost completely pointless without
758 // ClUseGlobalsGC (because then it only works on modules without
759 // globals, which are rare); it is a prerequisite for ClUseGlobalsGC;
760 // and both suffer from gold PR19002 for which UseGlobalsGC constructor
761 // argument is designed as workaround. Therefore, disable both
762 // ClWithComdat and ClUseGlobalsGC unless the frontend says it's ok to
763 // do globals-gc.
764 UseCtorComdat(UseGlobalsGC && ClWithComdat && !this->CompileKernel) {
765 C = &(M.getContext());
766 int LongSize = M.getDataLayout().getPointerSizeInBits();
767 IntptrTy = Type::getIntNTy(*C, LongSize);
768 TargetTriple = Triple(M.getTargetTriple());
769 Mapping = getShadowMapping(TargetTriple, LongSize, this->CompileKernel);
770 }
771
772 bool instrumentModule(Module &);
773
774 private:
775 void initializeCallbacks(Module &M);
776
777 bool InstrumentGlobals(IRBuilder<> &IRB, Module &M, bool *CtorComdat);
778 void InstrumentGlobalsCOFF(IRBuilder<> &IRB, Module &M,
779 ArrayRef<GlobalVariable *> ExtendedGlobals,
780 ArrayRef<Constant *> MetadataInitializers);
781 void InstrumentGlobalsELF(IRBuilder<> &IRB, Module &M,
782 ArrayRef<GlobalVariable *> ExtendedGlobals,
783 ArrayRef<Constant *> MetadataInitializers,
784 const std::string &UniqueModuleId);
785 void InstrumentGlobalsMachO(IRBuilder<> &IRB, Module &M,
786 ArrayRef<GlobalVariable *> ExtendedGlobals,
787 ArrayRef<Constant *> MetadataInitializers);
788 void
789 InstrumentGlobalsWithMetadataArray(IRBuilder<> &IRB, Module &M,
790 ArrayRef<GlobalVariable *> ExtendedGlobals,
791 ArrayRef<Constant *> MetadataInitializers);
792
793 GlobalVariable *CreateMetadataGlobal(Module &M, Constant *Initializer,
794 StringRef OriginalName);
795 void SetComdatForGlobalMetadata(GlobalVariable *G, GlobalVariable *Metadata,
796 StringRef InternalSuffix);
797 Instruction *CreateAsanModuleDtor(Module &M);
798
799 const GlobalVariable *getExcludedAliasedGlobal(const GlobalAlias &GA) const;
800 bool shouldInstrumentGlobal(GlobalVariable *G) const;
801 bool ShouldUseMachOGlobalsSection() const;
802 StringRef getGlobalMetadataSection() const;
803 void poisonOneInitializer(Function &GlobalInit, GlobalValue *ModuleName);
804 void createInitializerPoisonCalls(Module &M, GlobalValue *ModuleName);
getMinRedzoneSizeForGlobal() const805 uint64_t getMinRedzoneSizeForGlobal() const {
806 return getRedzoneSizeForScale(Mapping.Scale);
807 }
808 uint64_t getRedzoneSizeForGlobal(uint64_t SizeInBytes) const;
809 int GetAsanVersion(const Module &M) const;
810
811 const GlobalsMetadata &GlobalsMD;
812 bool CompileKernel;
813 bool Recover;
814 bool UseGlobalsGC;
815 bool UsePrivateAlias;
816 bool UseOdrIndicator;
817 bool UseCtorComdat;
818 Type *IntptrTy;
819 LLVMContext *C;
820 Triple TargetTriple;
821 ShadowMapping Mapping;
822 FunctionCallee AsanPoisonGlobals;
823 FunctionCallee AsanUnpoisonGlobals;
824 FunctionCallee AsanRegisterGlobals;
825 FunctionCallee AsanUnregisterGlobals;
826 FunctionCallee AsanRegisterImageGlobals;
827 FunctionCallee AsanUnregisterImageGlobals;
828 FunctionCallee AsanRegisterElfGlobals;
829 FunctionCallee AsanUnregisterElfGlobals;
830
831 Function *AsanCtorFunction = nullptr;
832 Function *AsanDtorFunction = nullptr;
833 };
834
835 class ModuleAddressSanitizerLegacyPass : public ModulePass {
836 public:
837 static char ID;
838
ModuleAddressSanitizerLegacyPass(bool CompileKernel=false,bool Recover=false,bool UseGlobalGC=true,bool UseOdrIndicator=false)839 explicit ModuleAddressSanitizerLegacyPass(bool CompileKernel = false,
840 bool Recover = false,
841 bool UseGlobalGC = true,
842 bool UseOdrIndicator = false)
843 : ModulePass(ID), CompileKernel(CompileKernel), Recover(Recover),
844 UseGlobalGC(UseGlobalGC), UseOdrIndicator(UseOdrIndicator) {
845 initializeModuleAddressSanitizerLegacyPassPass(
846 *PassRegistry::getPassRegistry());
847 }
848
getPassName() const849 StringRef getPassName() const override { return "ModuleAddressSanitizer"; }
850
getAnalysisUsage(AnalysisUsage & AU) const851 void getAnalysisUsage(AnalysisUsage &AU) const override {
852 AU.addRequired<ASanGlobalsMetadataWrapperPass>();
853 }
854
runOnModule(Module & M)855 bool runOnModule(Module &M) override {
856 GlobalsMetadata &GlobalsMD =
857 getAnalysis<ASanGlobalsMetadataWrapperPass>().getGlobalsMD();
858 ModuleAddressSanitizer ASanModule(M, &GlobalsMD, CompileKernel, Recover,
859 UseGlobalGC, UseOdrIndicator);
860 return ASanModule.instrumentModule(M);
861 }
862
863 private:
864 bool CompileKernel;
865 bool Recover;
866 bool UseGlobalGC;
867 bool UseOdrIndicator;
868 };
869
870 // Stack poisoning does not play well with exception handling.
871 // When an exception is thrown, we essentially bypass the code
872 // that unpoisones the stack. This is why the run-time library has
873 // to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
874 // stack in the interceptor. This however does not work inside the
875 // actual function which catches the exception. Most likely because the
876 // compiler hoists the load of the shadow value somewhere too high.
877 // This causes asan to report a non-existing bug on 453.povray.
878 // It sounds like an LLVM bug.
879 struct FunctionStackPoisoner : public InstVisitor<FunctionStackPoisoner> {
880 Function &F;
881 AddressSanitizer &ASan;
882 DIBuilder DIB;
883 LLVMContext *C;
884 Type *IntptrTy;
885 Type *IntptrPtrTy;
886 ShadowMapping Mapping;
887
888 SmallVector<AllocaInst *, 16> AllocaVec;
889 SmallVector<AllocaInst *, 16> StaticAllocasToMoveUp;
890 SmallVector<Instruction *, 8> RetVec;
891 unsigned StackAlignment;
892
893 FunctionCallee AsanStackMallocFunc[kMaxAsanStackMallocSizeClass + 1],
894 AsanStackFreeFunc[kMaxAsanStackMallocSizeClass + 1];
895 FunctionCallee AsanSetShadowFunc[0x100] = {};
896 FunctionCallee AsanPoisonStackMemoryFunc, AsanUnpoisonStackMemoryFunc;
897 FunctionCallee AsanAllocaPoisonFunc, AsanAllocasUnpoisonFunc;
898
899 // Stores a place and arguments of poisoning/unpoisoning call for alloca.
900 struct AllocaPoisonCall {
901 IntrinsicInst *InsBefore;
902 AllocaInst *AI;
903 uint64_t Size;
904 bool DoPoison;
905 };
906 SmallVector<AllocaPoisonCall, 8> DynamicAllocaPoisonCallVec;
907 SmallVector<AllocaPoisonCall, 8> StaticAllocaPoisonCallVec;
908 bool HasUntracedLifetimeIntrinsic = false;
909
910 SmallVector<AllocaInst *, 1> DynamicAllocaVec;
911 SmallVector<IntrinsicInst *, 1> StackRestoreVec;
912 AllocaInst *DynamicAllocaLayout = nullptr;
913 IntrinsicInst *LocalEscapeCall = nullptr;
914
915 bool HasInlineAsm = false;
916 bool HasReturnsTwiceCall = false;
917
FunctionStackPoisoner__anonfe46d5c80211::FunctionStackPoisoner918 FunctionStackPoisoner(Function &F, AddressSanitizer &ASan)
919 : F(F), ASan(ASan), DIB(*F.getParent(), /*AllowUnresolved*/ false),
920 C(ASan.C), IntptrTy(ASan.IntptrTy),
921 IntptrPtrTy(PointerType::get(IntptrTy, 0)), Mapping(ASan.Mapping),
922 StackAlignment(1 << Mapping.Scale) {}
923
runOnFunction__anonfe46d5c80211::FunctionStackPoisoner924 bool runOnFunction() {
925 if (!ClStack) return false;
926
927 if (ClRedzoneByvalArgs)
928 copyArgsPassedByValToAllocas();
929
930 // Collect alloca, ret, lifetime instructions etc.
931 for (BasicBlock *BB : depth_first(&F.getEntryBlock())) visit(*BB);
932
933 if (AllocaVec.empty() && DynamicAllocaVec.empty()) return false;
934
935 initializeCallbacks(*F.getParent());
936
937 if (HasUntracedLifetimeIntrinsic) {
938 // If there are lifetime intrinsics which couldn't be traced back to an
939 // alloca, we may not know exactly when a variable enters scope, and
940 // therefore should "fail safe" by not poisoning them.
941 StaticAllocaPoisonCallVec.clear();
942 DynamicAllocaPoisonCallVec.clear();
943 }
944
945 processDynamicAllocas();
946 processStaticAllocas();
947
948 if (ClDebugStack) {
949 LLVM_DEBUG(dbgs() << F);
950 }
951 return true;
952 }
953
954 // Arguments marked with the "byval" attribute are implicitly copied without
955 // using an alloca instruction. To produce redzones for those arguments, we
956 // copy them a second time into memory allocated with an alloca instruction.
957 void copyArgsPassedByValToAllocas();
958
959 // Finds all Alloca instructions and puts
960 // poisoned red zones around all of them.
961 // Then unpoison everything back before the function returns.
962 void processStaticAllocas();
963 void processDynamicAllocas();
964
965 void createDynamicAllocasInitStorage();
966
967 // ----------------------- Visitors.
968 /// Collect all Ret instructions, or the musttail call instruction if it
969 /// precedes the return instruction.
visitReturnInst__anonfe46d5c80211::FunctionStackPoisoner970 void visitReturnInst(ReturnInst &RI) {
971 if (CallInst *CI = RI.getParent()->getTerminatingMustTailCall())
972 RetVec.push_back(CI);
973 else
974 RetVec.push_back(&RI);
975 }
976
977 /// Collect all Resume instructions.
visitResumeInst__anonfe46d5c80211::FunctionStackPoisoner978 void visitResumeInst(ResumeInst &RI) { RetVec.push_back(&RI); }
979
980 /// Collect all CatchReturnInst instructions.
visitCleanupReturnInst__anonfe46d5c80211::FunctionStackPoisoner981 void visitCleanupReturnInst(CleanupReturnInst &CRI) { RetVec.push_back(&CRI); }
982
unpoisonDynamicAllocasBeforeInst__anonfe46d5c80211::FunctionStackPoisoner983 void unpoisonDynamicAllocasBeforeInst(Instruction *InstBefore,
984 Value *SavedStack) {
985 IRBuilder<> IRB(InstBefore);
986 Value *DynamicAreaPtr = IRB.CreatePtrToInt(SavedStack, IntptrTy);
987 // When we insert _asan_allocas_unpoison before @llvm.stackrestore, we
988 // need to adjust extracted SP to compute the address of the most recent
989 // alloca. We have a special @llvm.get.dynamic.area.offset intrinsic for
990 // this purpose.
991 if (!isa<ReturnInst>(InstBefore)) {
992 Function *DynamicAreaOffsetFunc = Intrinsic::getDeclaration(
993 InstBefore->getModule(), Intrinsic::get_dynamic_area_offset,
994 {IntptrTy});
995
996 Value *DynamicAreaOffset = IRB.CreateCall(DynamicAreaOffsetFunc, {});
997
998 DynamicAreaPtr = IRB.CreateAdd(IRB.CreatePtrToInt(SavedStack, IntptrTy),
999 DynamicAreaOffset);
1000 }
1001
1002 IRB.CreateCall(
1003 AsanAllocasUnpoisonFunc,
1004 {IRB.CreateLoad(IntptrTy, DynamicAllocaLayout), DynamicAreaPtr});
1005 }
1006
1007 // Unpoison dynamic allocas redzones.
unpoisonDynamicAllocas__anonfe46d5c80211::FunctionStackPoisoner1008 void unpoisonDynamicAllocas() {
1009 for (Instruction *Ret : RetVec)
1010 unpoisonDynamicAllocasBeforeInst(Ret, DynamicAllocaLayout);
1011
1012 for (Instruction *StackRestoreInst : StackRestoreVec)
1013 unpoisonDynamicAllocasBeforeInst(StackRestoreInst,
1014 StackRestoreInst->getOperand(0));
1015 }
1016
1017 // Deploy and poison redzones around dynamic alloca call. To do this, we
1018 // should replace this call with another one with changed parameters and
1019 // replace all its uses with new address, so
1020 // addr = alloca type, old_size, align
1021 // is replaced by
1022 // new_size = (old_size + additional_size) * sizeof(type)
1023 // tmp = alloca i8, new_size, max(align, 32)
1024 // addr = tmp + 32 (first 32 bytes are for the left redzone).
1025 // Additional_size is added to make new memory allocation contain not only
1026 // requested memory, but also left, partial and right redzones.
1027 void handleDynamicAllocaCall(AllocaInst *AI);
1028
1029 /// Collect Alloca instructions we want (and can) handle.
visitAllocaInst__anonfe46d5c80211::FunctionStackPoisoner1030 void visitAllocaInst(AllocaInst &AI) {
1031 if (!ASan.isInterestingAlloca(AI)) {
1032 if (AI.isStaticAlloca()) {
1033 // Skip over allocas that are present *before* the first instrumented
1034 // alloca, we don't want to move those around.
1035 if (AllocaVec.empty())
1036 return;
1037
1038 StaticAllocasToMoveUp.push_back(&AI);
1039 }
1040 return;
1041 }
1042
1043 StackAlignment = std::max(StackAlignment, AI.getAlignment());
1044 if (!AI.isStaticAlloca())
1045 DynamicAllocaVec.push_back(&AI);
1046 else
1047 AllocaVec.push_back(&AI);
1048 }
1049
1050 /// Collect lifetime intrinsic calls to check for use-after-scope
1051 /// errors.
visitIntrinsicInst__anonfe46d5c80211::FunctionStackPoisoner1052 void visitIntrinsicInst(IntrinsicInst &II) {
1053 Intrinsic::ID ID = II.getIntrinsicID();
1054 if (ID == Intrinsic::stackrestore) StackRestoreVec.push_back(&II);
1055 if (ID == Intrinsic::localescape) LocalEscapeCall = &II;
1056 if (!ASan.UseAfterScope)
1057 return;
1058 if (!II.isLifetimeStartOrEnd())
1059 return;
1060 // Found lifetime intrinsic, add ASan instrumentation if necessary.
1061 auto *Size = cast<ConstantInt>(II.getArgOperand(0));
1062 // If size argument is undefined, don't do anything.
1063 if (Size->isMinusOne()) return;
1064 // Check that size doesn't saturate uint64_t and can
1065 // be stored in IntptrTy.
1066 const uint64_t SizeValue = Size->getValue().getLimitedValue();
1067 if (SizeValue == ~0ULL ||
1068 !ConstantInt::isValueValidForType(IntptrTy, SizeValue))
1069 return;
1070 // Find alloca instruction that corresponds to llvm.lifetime argument.
1071 // Currently we can only handle lifetime markers pointing to the
1072 // beginning of the alloca.
1073 AllocaInst *AI = findAllocaForValue(II.getArgOperand(1), true);
1074 if (!AI) {
1075 HasUntracedLifetimeIntrinsic = true;
1076 return;
1077 }
1078 // We're interested only in allocas we can handle.
1079 if (!ASan.isInterestingAlloca(*AI))
1080 return;
1081 bool DoPoison = (ID == Intrinsic::lifetime_end);
1082 AllocaPoisonCall APC = {&II, AI, SizeValue, DoPoison};
1083 if (AI->isStaticAlloca())
1084 StaticAllocaPoisonCallVec.push_back(APC);
1085 else if (ClInstrumentDynamicAllocas)
1086 DynamicAllocaPoisonCallVec.push_back(APC);
1087 }
1088
visitCallBase__anonfe46d5c80211::FunctionStackPoisoner1089 void visitCallBase(CallBase &CB) {
1090 if (CallInst *CI = dyn_cast<CallInst>(&CB)) {
1091 HasInlineAsm |= CI->isInlineAsm() && &CB != ASan.LocalDynamicShadow;
1092 HasReturnsTwiceCall |= CI->canReturnTwice();
1093 }
1094 }
1095
1096 // ---------------------- Helpers.
1097 void initializeCallbacks(Module &M);
1098
1099 // Copies bytes from ShadowBytes into shadow memory for indexes where
1100 // ShadowMask is not zero. If ShadowMask[i] is zero, we assume that
1101 // ShadowBytes[i] is constantly zero and doesn't need to be overwritten.
1102 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1103 IRBuilder<> &IRB, Value *ShadowBase);
1104 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1105 size_t Begin, size_t End, IRBuilder<> &IRB,
1106 Value *ShadowBase);
1107 void copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
1108 ArrayRef<uint8_t> ShadowBytes, size_t Begin,
1109 size_t End, IRBuilder<> &IRB, Value *ShadowBase);
1110
1111 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> &IRB, bool DoPoison);
1112
1113 Value *createAllocaForLayout(IRBuilder<> &IRB, const ASanStackFrameLayout &L,
1114 bool Dynamic);
1115 PHINode *createPHI(IRBuilder<> &IRB, Value *Cond, Value *ValueIfTrue,
1116 Instruction *ThenTerm, Value *ValueIfFalse);
1117 };
1118
1119 } // end anonymous namespace
1120
parse(MDNode * MDN)1121 void LocationMetadata::parse(MDNode *MDN) {
1122 assert(MDN->getNumOperands() == 3);
1123 MDString *DIFilename = cast<MDString>(MDN->getOperand(0));
1124 Filename = DIFilename->getString();
1125 LineNo = mdconst::extract<ConstantInt>(MDN->getOperand(1))->getLimitedValue();
1126 ColumnNo =
1127 mdconst::extract<ConstantInt>(MDN->getOperand(2))->getLimitedValue();
1128 }
1129
1130 // FIXME: It would be cleaner to instead attach relevant metadata to the globals
1131 // we want to sanitize instead and reading this metadata on each pass over a
1132 // function instead of reading module level metadata at first.
GlobalsMetadata(Module & M)1133 GlobalsMetadata::GlobalsMetadata(Module &M) {
1134 NamedMDNode *Globals = M.getNamedMetadata("llvm.asan.globals");
1135 if (!Globals)
1136 return;
1137 for (auto MDN : Globals->operands()) {
1138 // Metadata node contains the global and the fields of "Entry".
1139 assert(MDN->getNumOperands() == 5);
1140 auto *V = mdconst::extract_or_null<Constant>(MDN->getOperand(0));
1141 // The optimizer may optimize away a global entirely.
1142 if (!V)
1143 continue;
1144 auto *StrippedV = V->stripPointerCasts();
1145 auto *GV = dyn_cast<GlobalVariable>(StrippedV);
1146 if (!GV)
1147 continue;
1148 // We can already have an entry for GV if it was merged with another
1149 // global.
1150 Entry &E = Entries[GV];
1151 if (auto *Loc = cast_or_null<MDNode>(MDN->getOperand(1)))
1152 E.SourceLoc.parse(Loc);
1153 if (auto *Name = cast_or_null<MDString>(MDN->getOperand(2)))
1154 E.Name = Name->getString();
1155 ConstantInt *IsDynInit = mdconst::extract<ConstantInt>(MDN->getOperand(3));
1156 E.IsDynInit |= IsDynInit->isOne();
1157 ConstantInt *IsExcluded =
1158 mdconst::extract<ConstantInt>(MDN->getOperand(4));
1159 E.IsExcluded |= IsExcluded->isOne();
1160 }
1161 }
1162
1163 AnalysisKey ASanGlobalsMetadataAnalysis::Key;
1164
run(Module & M,ModuleAnalysisManager & AM)1165 GlobalsMetadata ASanGlobalsMetadataAnalysis::run(Module &M,
1166 ModuleAnalysisManager &AM) {
1167 return GlobalsMetadata(M);
1168 }
1169
AddressSanitizerPass(bool CompileKernel,bool Recover,bool UseAfterScope)1170 AddressSanitizerPass::AddressSanitizerPass(bool CompileKernel, bool Recover,
1171 bool UseAfterScope)
1172 : CompileKernel(CompileKernel), Recover(Recover),
1173 UseAfterScope(UseAfterScope) {}
1174
run(Function & F,AnalysisManager<Function> & AM)1175 PreservedAnalyses AddressSanitizerPass::run(Function &F,
1176 AnalysisManager<Function> &AM) {
1177 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F);
1178 Module &M = *F.getParent();
1179 if (auto *R = MAMProxy.getCachedResult<ASanGlobalsMetadataAnalysis>(M)) {
1180 const TargetLibraryInfo *TLI = &AM.getResult<TargetLibraryAnalysis>(F);
1181 AddressSanitizer Sanitizer(M, R, CompileKernel, Recover, UseAfterScope);
1182 if (Sanitizer.instrumentFunction(F, TLI))
1183 return PreservedAnalyses::none();
1184 return PreservedAnalyses::all();
1185 }
1186
1187 report_fatal_error(
1188 "The ASanGlobalsMetadataAnalysis is required to run before "
1189 "AddressSanitizer can run");
1190 return PreservedAnalyses::all();
1191 }
1192
ModuleAddressSanitizerPass(bool CompileKernel,bool Recover,bool UseGlobalGC,bool UseOdrIndicator)1193 ModuleAddressSanitizerPass::ModuleAddressSanitizerPass(bool CompileKernel,
1194 bool Recover,
1195 bool UseGlobalGC,
1196 bool UseOdrIndicator)
1197 : CompileKernel(CompileKernel), Recover(Recover), UseGlobalGC(UseGlobalGC),
1198 UseOdrIndicator(UseOdrIndicator) {}
1199
run(Module & M,AnalysisManager<Module> & AM)1200 PreservedAnalyses ModuleAddressSanitizerPass::run(Module &M,
1201 AnalysisManager<Module> &AM) {
1202 GlobalsMetadata &GlobalsMD = AM.getResult<ASanGlobalsMetadataAnalysis>(M);
1203 ModuleAddressSanitizer Sanitizer(M, &GlobalsMD, CompileKernel, Recover,
1204 UseGlobalGC, UseOdrIndicator);
1205 if (Sanitizer.instrumentModule(M))
1206 return PreservedAnalyses::none();
1207 return PreservedAnalyses::all();
1208 }
1209
1210 INITIALIZE_PASS(ASanGlobalsMetadataWrapperPass, "asan-globals-md",
1211 "Read metadata to mark which globals should be instrumented "
1212 "when running ASan.",
1213 false, true)
1214
1215 char AddressSanitizerLegacyPass::ID = 0;
1216
1217 INITIALIZE_PASS_BEGIN(
1218 AddressSanitizerLegacyPass, "asan",
1219 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.", false,
1220 false)
INITIALIZE_PASS_DEPENDENCY(ASanGlobalsMetadataWrapperPass)1221 INITIALIZE_PASS_DEPENDENCY(ASanGlobalsMetadataWrapperPass)
1222 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1223 INITIALIZE_PASS_END(
1224 AddressSanitizerLegacyPass, "asan",
1225 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.", false,
1226 false)
1227
1228 FunctionPass *llvm::createAddressSanitizerFunctionPass(bool CompileKernel,
1229 bool Recover,
1230 bool UseAfterScope) {
1231 assert(!CompileKernel || Recover);
1232 return new AddressSanitizerLegacyPass(CompileKernel, Recover, UseAfterScope);
1233 }
1234
1235 char ModuleAddressSanitizerLegacyPass::ID = 0;
1236
1237 INITIALIZE_PASS(
1238 ModuleAddressSanitizerLegacyPass, "asan-module",
1239 "AddressSanitizer: detects use-after-free and out-of-bounds bugs."
1240 "ModulePass",
1241 false, false)
1242
createModuleAddressSanitizerLegacyPassPass(bool CompileKernel,bool Recover,bool UseGlobalsGC,bool UseOdrIndicator)1243 ModulePass *llvm::createModuleAddressSanitizerLegacyPassPass(
1244 bool CompileKernel, bool Recover, bool UseGlobalsGC, bool UseOdrIndicator) {
1245 assert(!CompileKernel || Recover);
1246 return new ModuleAddressSanitizerLegacyPass(CompileKernel, Recover,
1247 UseGlobalsGC, UseOdrIndicator);
1248 }
1249
TypeSizeToSizeIndex(uint32_t TypeSize)1250 static size_t TypeSizeToSizeIndex(uint32_t TypeSize) {
1251 size_t Res = countTrailingZeros(TypeSize / 8);
1252 assert(Res < kNumberOfAccessSizes);
1253 return Res;
1254 }
1255
1256 /// Create a global describing a source location.
createPrivateGlobalForSourceLoc(Module & M,LocationMetadata MD)1257 static GlobalVariable *createPrivateGlobalForSourceLoc(Module &M,
1258 LocationMetadata MD) {
1259 Constant *LocData[] = {
1260 createPrivateGlobalForString(M, MD.Filename, true, kAsanGenPrefix),
1261 ConstantInt::get(Type::getInt32Ty(M.getContext()), MD.LineNo),
1262 ConstantInt::get(Type::getInt32Ty(M.getContext()), MD.ColumnNo),
1263 };
1264 auto LocStruct = ConstantStruct::getAnon(LocData);
1265 auto GV = new GlobalVariable(M, LocStruct->getType(), true,
1266 GlobalValue::PrivateLinkage, LocStruct,
1267 kAsanGenPrefix);
1268 GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
1269 return GV;
1270 }
1271
1272 /// Check if \p G has been created by a trusted compiler pass.
GlobalWasGeneratedByCompiler(GlobalVariable * G)1273 static bool GlobalWasGeneratedByCompiler(GlobalVariable *G) {
1274 // Do not instrument @llvm.global_ctors, @llvm.used, etc.
1275 if (G->getName().startswith("llvm."))
1276 return true;
1277
1278 // Do not instrument asan globals.
1279 if (G->getName().startswith(kAsanGenPrefix) ||
1280 G->getName().startswith(kSanCovGenPrefix) ||
1281 G->getName().startswith(kODRGenPrefix))
1282 return true;
1283
1284 // Do not instrument gcov counter arrays.
1285 if (G->getName() == "__llvm_gcov_ctr")
1286 return true;
1287
1288 return false;
1289 }
1290
memToShadow(Value * Shadow,IRBuilder<> & IRB)1291 Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
1292 // Shadow >> scale
1293 Shadow = IRB.CreateLShr(Shadow, Mapping.Scale);
1294 if (Mapping.Offset == 0) return Shadow;
1295 // (Shadow >> scale) | offset
1296 Value *ShadowBase;
1297 if (LocalDynamicShadow)
1298 ShadowBase = LocalDynamicShadow;
1299 else
1300 ShadowBase = ConstantInt::get(IntptrTy, Mapping.Offset);
1301 if (Mapping.OrShadowOffset)
1302 return IRB.CreateOr(Shadow, ShadowBase);
1303 else
1304 return IRB.CreateAdd(Shadow, ShadowBase);
1305 }
1306
1307 // Instrument memset/memmove/memcpy
instrumentMemIntrinsic(MemIntrinsic * MI)1308 void AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) {
1309 IRBuilder<> IRB(MI);
1310 if (isa<MemTransferInst>(MI)) {
1311 IRB.CreateCall(
1312 isa<MemMoveInst>(MI) ? AsanMemmove : AsanMemcpy,
1313 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()),
1314 IRB.CreatePointerCast(MI->getOperand(1), IRB.getInt8PtrTy()),
1315 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)});
1316 } else if (isa<MemSetInst>(MI)) {
1317 IRB.CreateCall(
1318 AsanMemset,
1319 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()),
1320 IRB.CreateIntCast(MI->getOperand(1), IRB.getInt32Ty(), false),
1321 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)});
1322 }
1323 MI->eraseFromParent();
1324 }
1325
1326 /// Check if we want (and can) handle this alloca.
isInterestingAlloca(const AllocaInst & AI)1327 bool AddressSanitizer::isInterestingAlloca(const AllocaInst &AI) {
1328 auto PreviouslySeenAllocaInfo = ProcessedAllocas.find(&AI);
1329
1330 if (PreviouslySeenAllocaInfo != ProcessedAllocas.end())
1331 return PreviouslySeenAllocaInfo->getSecond();
1332
1333 bool IsInteresting =
1334 (AI.getAllocatedType()->isSized() &&
1335 // alloca() may be called with 0 size, ignore it.
1336 ((!AI.isStaticAlloca()) || getAllocaSizeInBytes(AI) > 0) &&
1337 // We are only interested in allocas not promotable to registers.
1338 // Promotable allocas are common under -O0.
1339 (!ClSkipPromotableAllocas || !isAllocaPromotable(&AI)) &&
1340 // inalloca allocas are not treated as static, and we don't want
1341 // dynamic alloca instrumentation for them as well.
1342 !AI.isUsedWithInAlloca() &&
1343 // swifterror allocas are register promoted by ISel
1344 !AI.isSwiftError());
1345
1346 ProcessedAllocas[&AI] = IsInteresting;
1347 return IsInteresting;
1348 }
1349
ignoreAccess(Value * Ptr)1350 bool AddressSanitizer::ignoreAccess(Value *Ptr) {
1351 // Do not instrument acesses from different address spaces; we cannot deal
1352 // with them.
1353 Type *PtrTy = cast<PointerType>(Ptr->getType()->getScalarType());
1354 if (PtrTy->getPointerAddressSpace() != 0)
1355 return true;
1356
1357 // Ignore swifterror addresses.
1358 // swifterror memory addresses are mem2reg promoted by instruction
1359 // selection. As such they cannot have regular uses like an instrumentation
1360 // function and it makes no sense to track them as memory.
1361 if (Ptr->isSwiftError())
1362 return true;
1363
1364 // Treat memory accesses to promotable allocas as non-interesting since they
1365 // will not cause memory violations. This greatly speeds up the instrumented
1366 // executable at -O0.
1367 if (auto AI = dyn_cast_or_null<AllocaInst>(Ptr))
1368 if (ClSkipPromotableAllocas && !isInterestingAlloca(*AI))
1369 return true;
1370
1371 return false;
1372 }
1373
getInterestingMemoryOperands(Instruction * I,SmallVectorImpl<InterestingMemoryOperand> & Interesting)1374 void AddressSanitizer::getInterestingMemoryOperands(
1375 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting) {
1376 // Skip memory accesses inserted by another instrumentation.
1377 if (I->hasMetadata("nosanitize"))
1378 return;
1379
1380 // Do not instrument the load fetching the dynamic shadow address.
1381 if (LocalDynamicShadow == I)
1382 return;
1383
1384 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1385 if (!ClInstrumentReads || ignoreAccess(LI->getPointerOperand()))
1386 return;
1387 Interesting.emplace_back(I, LI->getPointerOperandIndex(), false,
1388 LI->getType(), LI->getAlign());
1389 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1390 if (!ClInstrumentWrites || ignoreAccess(SI->getPointerOperand()))
1391 return;
1392 Interesting.emplace_back(I, SI->getPointerOperandIndex(), true,
1393 SI->getValueOperand()->getType(), SI->getAlign());
1394 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
1395 if (!ClInstrumentAtomics || ignoreAccess(RMW->getPointerOperand()))
1396 return;
1397 Interesting.emplace_back(I, RMW->getPointerOperandIndex(), true,
1398 RMW->getValOperand()->getType(), None);
1399 } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) {
1400 if (!ClInstrumentAtomics || ignoreAccess(XCHG->getPointerOperand()))
1401 return;
1402 Interesting.emplace_back(I, XCHG->getPointerOperandIndex(), true,
1403 XCHG->getCompareOperand()->getType(), None);
1404 } else if (auto CI = dyn_cast<CallInst>(I)) {
1405 auto *F = CI->getCalledFunction();
1406 if (F && (F->getName().startswith("llvm.masked.load.") ||
1407 F->getName().startswith("llvm.masked.store."))) {
1408 bool IsWrite = F->getName().startswith("llvm.masked.store.");
1409 // Masked store has an initial operand for the value.
1410 unsigned OpOffset = IsWrite ? 1 : 0;
1411 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1412 return;
1413
1414 auto BasePtr = CI->getOperand(OpOffset);
1415 if (ignoreAccess(BasePtr))
1416 return;
1417 auto Ty = cast<PointerType>(BasePtr->getType())->getElementType();
1418 MaybeAlign Alignment = Align(1);
1419 // Otherwise no alignment guarantees. We probably got Undef.
1420 if (auto *Op = dyn_cast<ConstantInt>(CI->getOperand(1 + OpOffset)))
1421 Alignment = Op->getMaybeAlignValue();
1422 Value *Mask = CI->getOperand(2 + OpOffset);
1423 Interesting.emplace_back(I, OpOffset, IsWrite, Ty, Alignment, Mask);
1424 } else {
1425 for (unsigned ArgNo = 0; ArgNo < CI->getNumArgOperands(); ArgNo++) {
1426 if (!ClInstrumentByval || !CI->isByValArgument(ArgNo) ||
1427 ignoreAccess(CI->getArgOperand(ArgNo)))
1428 continue;
1429 Type *Ty = CI->getParamByValType(ArgNo);
1430 Interesting.emplace_back(I, ArgNo, false, Ty, Align(1));
1431 }
1432 }
1433 }
1434 }
1435
isPointerOperand(Value * V)1436 static bool isPointerOperand(Value *V) {
1437 return V->getType()->isPointerTy() || isa<PtrToIntInst>(V);
1438 }
1439
1440 // This is a rough heuristic; it may cause both false positives and
1441 // false negatives. The proper implementation requires cooperation with
1442 // the frontend.
isInterestingPointerComparison(Instruction * I)1443 static bool isInterestingPointerComparison(Instruction *I) {
1444 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(I)) {
1445 if (!Cmp->isRelational())
1446 return false;
1447 } else {
1448 return false;
1449 }
1450 return isPointerOperand(I->getOperand(0)) &&
1451 isPointerOperand(I->getOperand(1));
1452 }
1453
1454 // This is a rough heuristic; it may cause both false positives and
1455 // false negatives. The proper implementation requires cooperation with
1456 // the frontend.
isInterestingPointerSubtraction(Instruction * I)1457 static bool isInterestingPointerSubtraction(Instruction *I) {
1458 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
1459 if (BO->getOpcode() != Instruction::Sub)
1460 return false;
1461 } else {
1462 return false;
1463 }
1464 return isPointerOperand(I->getOperand(0)) &&
1465 isPointerOperand(I->getOperand(1));
1466 }
1467
GlobalIsLinkerInitialized(GlobalVariable * G)1468 bool AddressSanitizer::GlobalIsLinkerInitialized(GlobalVariable *G) {
1469 // If a global variable does not have dynamic initialization we don't
1470 // have to instrument it. However, if a global does not have initializer
1471 // at all, we assume it has dynamic initializer (in other TU).
1472 //
1473 // FIXME: Metadata should be attched directly to the global directly instead
1474 // of being added to llvm.asan.globals.
1475 return G->hasInitializer() && !GlobalsMD.get(G).IsDynInit;
1476 }
1477
instrumentPointerComparisonOrSubtraction(Instruction * I)1478 void AddressSanitizer::instrumentPointerComparisonOrSubtraction(
1479 Instruction *I) {
1480 IRBuilder<> IRB(I);
1481 FunctionCallee F = isa<ICmpInst>(I) ? AsanPtrCmpFunction : AsanPtrSubFunction;
1482 Value *Param[2] = {I->getOperand(0), I->getOperand(1)};
1483 for (Value *&i : Param) {
1484 if (i->getType()->isPointerTy())
1485 i = IRB.CreatePointerCast(i, IntptrTy);
1486 }
1487 IRB.CreateCall(F, Param);
1488 }
1489
doInstrumentAddress(AddressSanitizer * Pass,Instruction * I,Instruction * InsertBefore,Value * Addr,MaybeAlign Alignment,unsigned Granularity,uint32_t TypeSize,bool IsWrite,Value * SizeArgument,bool UseCalls,uint32_t Exp)1490 static void doInstrumentAddress(AddressSanitizer *Pass, Instruction *I,
1491 Instruction *InsertBefore, Value *Addr,
1492 MaybeAlign Alignment, unsigned Granularity,
1493 uint32_t TypeSize, bool IsWrite,
1494 Value *SizeArgument, bool UseCalls,
1495 uint32_t Exp) {
1496 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check
1497 // if the data is properly aligned.
1498 if ((TypeSize == 8 || TypeSize == 16 || TypeSize == 32 || TypeSize == 64 ||
1499 TypeSize == 128) &&
1500 (!Alignment || *Alignment >= Granularity || *Alignment >= TypeSize / 8))
1501 return Pass->instrumentAddress(I, InsertBefore, Addr, TypeSize, IsWrite,
1502 nullptr, UseCalls, Exp);
1503 Pass->instrumentUnusualSizeOrAlignment(I, InsertBefore, Addr, TypeSize,
1504 IsWrite, nullptr, UseCalls, Exp);
1505 }
1506
instrumentMaskedLoadOrStore(AddressSanitizer * Pass,const DataLayout & DL,Type * IntptrTy,Value * Mask,Instruction * I,Value * Addr,MaybeAlign Alignment,unsigned Granularity,uint32_t TypeSize,bool IsWrite,Value * SizeArgument,bool UseCalls,uint32_t Exp)1507 static void instrumentMaskedLoadOrStore(AddressSanitizer *Pass,
1508 const DataLayout &DL, Type *IntptrTy,
1509 Value *Mask, Instruction *I,
1510 Value *Addr, MaybeAlign Alignment,
1511 unsigned Granularity, uint32_t TypeSize,
1512 bool IsWrite, Value *SizeArgument,
1513 bool UseCalls, uint32_t Exp) {
1514 auto *VTy = cast<FixedVectorType>(
1515 cast<PointerType>(Addr->getType())->getElementType());
1516 uint64_t ElemTypeSize = DL.getTypeStoreSizeInBits(VTy->getScalarType());
1517 unsigned Num = VTy->getNumElements();
1518 auto Zero = ConstantInt::get(IntptrTy, 0);
1519 for (unsigned Idx = 0; Idx < Num; ++Idx) {
1520 Value *InstrumentedAddress = nullptr;
1521 Instruction *InsertBefore = I;
1522 if (auto *Vector = dyn_cast<ConstantVector>(Mask)) {
1523 // dyn_cast as we might get UndefValue
1524 if (auto *Masked = dyn_cast<ConstantInt>(Vector->getOperand(Idx))) {
1525 if (Masked->isZero())
1526 // Mask is constant false, so no instrumentation needed.
1527 continue;
1528 // If we have a true or undef value, fall through to doInstrumentAddress
1529 // with InsertBefore == I
1530 }
1531 } else {
1532 IRBuilder<> IRB(I);
1533 Value *MaskElem = IRB.CreateExtractElement(Mask, Idx);
1534 Instruction *ThenTerm = SplitBlockAndInsertIfThen(MaskElem, I, false);
1535 InsertBefore = ThenTerm;
1536 }
1537
1538 IRBuilder<> IRB(InsertBefore);
1539 InstrumentedAddress =
1540 IRB.CreateGEP(VTy, Addr, {Zero, ConstantInt::get(IntptrTy, Idx)});
1541 doInstrumentAddress(Pass, I, InsertBefore, InstrumentedAddress, Alignment,
1542 Granularity, ElemTypeSize, IsWrite, SizeArgument,
1543 UseCalls, Exp);
1544 }
1545 }
1546
instrumentMop(ObjectSizeOffsetVisitor & ObjSizeVis,InterestingMemoryOperand & O,bool UseCalls,const DataLayout & DL)1547 void AddressSanitizer::instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
1548 InterestingMemoryOperand &O, bool UseCalls,
1549 const DataLayout &DL) {
1550 Value *Addr = O.getPtr();
1551
1552 // Optimization experiments.
1553 // The experiments can be used to evaluate potential optimizations that remove
1554 // instrumentation (assess false negatives). Instead of completely removing
1555 // some instrumentation, you set Exp to a non-zero value (mask of optimization
1556 // experiments that want to remove instrumentation of this instruction).
1557 // If Exp is non-zero, this pass will emit special calls into runtime
1558 // (e.g. __asan_report_exp_load1 instead of __asan_report_load1). These calls
1559 // make runtime terminate the program in a special way (with a different
1560 // exit status). Then you run the new compiler on a buggy corpus, collect
1561 // the special terminations (ideally, you don't see them at all -- no false
1562 // negatives) and make the decision on the optimization.
1563 uint32_t Exp = ClForceExperiment;
1564
1565 if (ClOpt && ClOptGlobals) {
1566 // If initialization order checking is disabled, a simple access to a
1567 // dynamically initialized global is always valid.
1568 GlobalVariable *G = dyn_cast<GlobalVariable>(getUnderlyingObject(Addr));
1569 if (G && (!ClInitializers || GlobalIsLinkerInitialized(G)) &&
1570 isSafeAccess(ObjSizeVis, Addr, O.TypeSize)) {
1571 NumOptimizedAccessesToGlobalVar++;
1572 return;
1573 }
1574 }
1575
1576 if (ClOpt && ClOptStack) {
1577 // A direct inbounds access to a stack variable is always valid.
1578 if (isa<AllocaInst>(getUnderlyingObject(Addr)) &&
1579 isSafeAccess(ObjSizeVis, Addr, O.TypeSize)) {
1580 NumOptimizedAccessesToStackVar++;
1581 return;
1582 }
1583 }
1584
1585 if (O.IsWrite)
1586 NumInstrumentedWrites++;
1587 else
1588 NumInstrumentedReads++;
1589
1590 unsigned Granularity = 1 << Mapping.Scale;
1591 if (O.MaybeMask) {
1592 instrumentMaskedLoadOrStore(this, DL, IntptrTy, O.MaybeMask, O.getInsn(),
1593 Addr, O.Alignment, Granularity, O.TypeSize,
1594 O.IsWrite, nullptr, UseCalls, Exp);
1595 } else {
1596 doInstrumentAddress(this, O.getInsn(), O.getInsn(), Addr, O.Alignment,
1597 Granularity, O.TypeSize, O.IsWrite, nullptr, UseCalls,
1598 Exp);
1599 }
1600 }
1601
generateCrashCode(Instruction * InsertBefore,Value * Addr,bool IsWrite,size_t AccessSizeIndex,Value * SizeArgument,uint32_t Exp)1602 Instruction *AddressSanitizer::generateCrashCode(Instruction *InsertBefore,
1603 Value *Addr, bool IsWrite,
1604 size_t AccessSizeIndex,
1605 Value *SizeArgument,
1606 uint32_t Exp) {
1607 IRBuilder<> IRB(InsertBefore);
1608 Value *ExpVal = Exp == 0 ? nullptr : ConstantInt::get(IRB.getInt32Ty(), Exp);
1609 CallInst *Call = nullptr;
1610 if (SizeArgument) {
1611 if (Exp == 0)
1612 Call = IRB.CreateCall(AsanErrorCallbackSized[IsWrite][0],
1613 {Addr, SizeArgument});
1614 else
1615 Call = IRB.CreateCall(AsanErrorCallbackSized[IsWrite][1],
1616 {Addr, SizeArgument, ExpVal});
1617 } else {
1618 if (Exp == 0)
1619 Call =
1620 IRB.CreateCall(AsanErrorCallback[IsWrite][0][AccessSizeIndex], Addr);
1621 else
1622 Call = IRB.CreateCall(AsanErrorCallback[IsWrite][1][AccessSizeIndex],
1623 {Addr, ExpVal});
1624 }
1625
1626 Call->setCannotMerge();
1627 return Call;
1628 }
1629
createSlowPathCmp(IRBuilder<> & IRB,Value * AddrLong,Value * ShadowValue,uint32_t TypeSize)1630 Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
1631 Value *ShadowValue,
1632 uint32_t TypeSize) {
1633 size_t Granularity = static_cast<size_t>(1) << Mapping.Scale;
1634 // Addr & (Granularity - 1)
1635 Value *LastAccessedByte =
1636 IRB.CreateAnd(AddrLong, ConstantInt::get(IntptrTy, Granularity - 1));
1637 // (Addr & (Granularity - 1)) + size - 1
1638 if (TypeSize / 8 > 1)
1639 LastAccessedByte = IRB.CreateAdd(
1640 LastAccessedByte, ConstantInt::get(IntptrTy, TypeSize / 8 - 1));
1641 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
1642 LastAccessedByte =
1643 IRB.CreateIntCast(LastAccessedByte, ShadowValue->getType(), false);
1644 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
1645 return IRB.CreateICmpSGE(LastAccessedByte, ShadowValue);
1646 }
1647
instrumentAddress(Instruction * OrigIns,Instruction * InsertBefore,Value * Addr,uint32_t TypeSize,bool IsWrite,Value * SizeArgument,bool UseCalls,uint32_t Exp)1648 void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
1649 Instruction *InsertBefore, Value *Addr,
1650 uint32_t TypeSize, bool IsWrite,
1651 Value *SizeArgument, bool UseCalls,
1652 uint32_t Exp) {
1653 bool IsMyriad = TargetTriple.getVendor() == llvm::Triple::Myriad;
1654
1655 IRBuilder<> IRB(InsertBefore);
1656 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
1657 size_t AccessSizeIndex = TypeSizeToSizeIndex(TypeSize);
1658
1659 if (UseCalls) {
1660 if (Exp == 0)
1661 IRB.CreateCall(AsanMemoryAccessCallback[IsWrite][0][AccessSizeIndex],
1662 AddrLong);
1663 else
1664 IRB.CreateCall(AsanMemoryAccessCallback[IsWrite][1][AccessSizeIndex],
1665 {AddrLong, ConstantInt::get(IRB.getInt32Ty(), Exp)});
1666 return;
1667 }
1668
1669 if (IsMyriad) {
1670 // Strip the cache bit and do range check.
1671 // AddrLong &= ~kMyriadCacheBitMask32
1672 AddrLong = IRB.CreateAnd(AddrLong, ~kMyriadCacheBitMask32);
1673 // Tag = AddrLong >> kMyriadTagShift
1674 Value *Tag = IRB.CreateLShr(AddrLong, kMyriadTagShift);
1675 // Tag == kMyriadDDRTag
1676 Value *TagCheck =
1677 IRB.CreateICmpEQ(Tag, ConstantInt::get(IntptrTy, kMyriadDDRTag));
1678
1679 Instruction *TagCheckTerm =
1680 SplitBlockAndInsertIfThen(TagCheck, InsertBefore, false,
1681 MDBuilder(*C).createBranchWeights(1, 100000));
1682 assert(cast<BranchInst>(TagCheckTerm)->isUnconditional());
1683 IRB.SetInsertPoint(TagCheckTerm);
1684 InsertBefore = TagCheckTerm;
1685 }
1686
1687 Type *ShadowTy =
1688 IntegerType::get(*C, std::max(8U, TypeSize >> Mapping.Scale));
1689 Type *ShadowPtrTy = PointerType::get(ShadowTy, 0);
1690 Value *ShadowPtr = memToShadow(AddrLong, IRB);
1691 Value *CmpVal = Constant::getNullValue(ShadowTy);
1692 Value *ShadowValue =
1693 IRB.CreateLoad(ShadowTy, IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy));
1694
1695 Value *Cmp = IRB.CreateICmpNE(ShadowValue, CmpVal);
1696 size_t Granularity = 1ULL << Mapping.Scale;
1697 Instruction *CrashTerm = nullptr;
1698
1699 if (ClAlwaysSlowPath || (TypeSize < 8 * Granularity)) {
1700 // We use branch weights for the slow path check, to indicate that the slow
1701 // path is rarely taken. This seems to be the case for SPEC benchmarks.
1702 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
1703 Cmp, InsertBefore, false, MDBuilder(*C).createBranchWeights(1, 100000));
1704 assert(cast<BranchInst>(CheckTerm)->isUnconditional());
1705 BasicBlock *NextBB = CheckTerm->getSuccessor(0);
1706 IRB.SetInsertPoint(CheckTerm);
1707 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeSize);
1708 if (Recover) {
1709 CrashTerm = SplitBlockAndInsertIfThen(Cmp2, CheckTerm, false);
1710 } else {
1711 BasicBlock *CrashBlock =
1712 BasicBlock::Create(*C, "", NextBB->getParent(), NextBB);
1713 CrashTerm = new UnreachableInst(*C, CrashBlock);
1714 BranchInst *NewTerm = BranchInst::Create(CrashBlock, NextBB, Cmp2);
1715 ReplaceInstWithInst(CheckTerm, NewTerm);
1716 }
1717 } else {
1718 CrashTerm = SplitBlockAndInsertIfThen(Cmp, InsertBefore, !Recover);
1719 }
1720
1721 Instruction *Crash = generateCrashCode(CrashTerm, AddrLong, IsWrite,
1722 AccessSizeIndex, SizeArgument, Exp);
1723 Crash->setDebugLoc(OrigIns->getDebugLoc());
1724 }
1725
1726 // Instrument unusual size or unusual alignment.
1727 // We can not do it with a single check, so we do 1-byte check for the first
1728 // and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
1729 // to report the actual access size.
instrumentUnusualSizeOrAlignment(Instruction * I,Instruction * InsertBefore,Value * Addr,uint32_t TypeSize,bool IsWrite,Value * SizeArgument,bool UseCalls,uint32_t Exp)1730 void AddressSanitizer::instrumentUnusualSizeOrAlignment(
1731 Instruction *I, Instruction *InsertBefore, Value *Addr, uint32_t TypeSize,
1732 bool IsWrite, Value *SizeArgument, bool UseCalls, uint32_t Exp) {
1733 IRBuilder<> IRB(InsertBefore);
1734 Value *Size = ConstantInt::get(IntptrTy, TypeSize / 8);
1735 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
1736 if (UseCalls) {
1737 if (Exp == 0)
1738 IRB.CreateCall(AsanMemoryAccessCallbackSized[IsWrite][0],
1739 {AddrLong, Size});
1740 else
1741 IRB.CreateCall(AsanMemoryAccessCallbackSized[IsWrite][1],
1742 {AddrLong, Size, ConstantInt::get(IRB.getInt32Ty(), Exp)});
1743 } else {
1744 Value *LastByte = IRB.CreateIntToPtr(
1745 IRB.CreateAdd(AddrLong, ConstantInt::get(IntptrTy, TypeSize / 8 - 1)),
1746 Addr->getType());
1747 instrumentAddress(I, InsertBefore, Addr, 8, IsWrite, Size, false, Exp);
1748 instrumentAddress(I, InsertBefore, LastByte, 8, IsWrite, Size, false, Exp);
1749 }
1750 }
1751
poisonOneInitializer(Function & GlobalInit,GlobalValue * ModuleName)1752 void ModuleAddressSanitizer::poisonOneInitializer(Function &GlobalInit,
1753 GlobalValue *ModuleName) {
1754 // Set up the arguments to our poison/unpoison functions.
1755 IRBuilder<> IRB(&GlobalInit.front(),
1756 GlobalInit.front().getFirstInsertionPt());
1757
1758 // Add a call to poison all external globals before the given function starts.
1759 Value *ModuleNameAddr = ConstantExpr::getPointerCast(ModuleName, IntptrTy);
1760 IRB.CreateCall(AsanPoisonGlobals, ModuleNameAddr);
1761
1762 // Add calls to unpoison all globals before each return instruction.
1763 for (auto &BB : GlobalInit.getBasicBlockList())
1764 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB.getTerminator()))
1765 CallInst::Create(AsanUnpoisonGlobals, "", RI);
1766 }
1767
createInitializerPoisonCalls(Module & M,GlobalValue * ModuleName)1768 void ModuleAddressSanitizer::createInitializerPoisonCalls(
1769 Module &M, GlobalValue *ModuleName) {
1770 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
1771 if (!GV)
1772 return;
1773
1774 ConstantArray *CA = dyn_cast<ConstantArray>(GV->getInitializer());
1775 if (!CA)
1776 return;
1777
1778 for (Use &OP : CA->operands()) {
1779 if (isa<ConstantAggregateZero>(OP)) continue;
1780 ConstantStruct *CS = cast<ConstantStruct>(OP);
1781
1782 // Must have a function or null ptr.
1783 if (Function *F = dyn_cast<Function>(CS->getOperand(1))) {
1784 if (F->getName() == kAsanModuleCtorName) continue;
1785 auto *Priority = cast<ConstantInt>(CS->getOperand(0));
1786 // Don't instrument CTORs that will run before asan.module_ctor.
1787 if (Priority->getLimitedValue() <= GetCtorAndDtorPriority(TargetTriple))
1788 continue;
1789 poisonOneInitializer(*F, ModuleName);
1790 }
1791 }
1792 }
1793
1794 const GlobalVariable *
getExcludedAliasedGlobal(const GlobalAlias & GA) const1795 ModuleAddressSanitizer::getExcludedAliasedGlobal(const GlobalAlias &GA) const {
1796 // In case this function should be expanded to include rules that do not just
1797 // apply when CompileKernel is true, either guard all existing rules with an
1798 // 'if (CompileKernel) { ... }' or be absolutely sure that all these rules
1799 // should also apply to user space.
1800 assert(CompileKernel && "Only expecting to be called when compiling kernel");
1801
1802 const Constant *C = GA.getAliasee();
1803
1804 // When compiling the kernel, globals that are aliased by symbols prefixed
1805 // by "__" are special and cannot be padded with a redzone.
1806 if (GA.getName().startswith("__"))
1807 return dyn_cast<GlobalVariable>(C->stripPointerCastsAndAliases());
1808
1809 return nullptr;
1810 }
1811
shouldInstrumentGlobal(GlobalVariable * G) const1812 bool ModuleAddressSanitizer::shouldInstrumentGlobal(GlobalVariable *G) const {
1813 Type *Ty = G->getValueType();
1814 LLVM_DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
1815
1816 // FIXME: Metadata should be attched directly to the global directly instead
1817 // of being added to llvm.asan.globals.
1818 if (GlobalsMD.get(G).IsExcluded) return false;
1819 if (!Ty->isSized()) return false;
1820 if (!G->hasInitializer()) return false;
1821 // Only instrument globals of default address spaces
1822 if (G->getAddressSpace()) return false;
1823 if (GlobalWasGeneratedByCompiler(G)) return false; // Our own globals.
1824 // Two problems with thread-locals:
1825 // - The address of the main thread's copy can't be computed at link-time.
1826 // - Need to poison all copies, not just the main thread's one.
1827 if (G->isThreadLocal()) return false;
1828 // For now, just ignore this Global if the alignment is large.
1829 if (G->getAlignment() > getMinRedzoneSizeForGlobal()) return false;
1830
1831 // For non-COFF targets, only instrument globals known to be defined by this
1832 // TU.
1833 // FIXME: We can instrument comdat globals on ELF if we are using the
1834 // GC-friendly metadata scheme.
1835 if (!TargetTriple.isOSBinFormatCOFF()) {
1836 if (!G->hasExactDefinition() || G->hasComdat())
1837 return false;
1838 } else {
1839 // On COFF, don't instrument non-ODR linkages.
1840 if (G->isInterposable())
1841 return false;
1842 }
1843
1844 // If a comdat is present, it must have a selection kind that implies ODR
1845 // semantics: no duplicates, any, or exact match.
1846 if (Comdat *C = G->getComdat()) {
1847 switch (C->getSelectionKind()) {
1848 case Comdat::Any:
1849 case Comdat::ExactMatch:
1850 case Comdat::NoDuplicates:
1851 break;
1852 case Comdat::Largest:
1853 case Comdat::SameSize:
1854 return false;
1855 }
1856 }
1857
1858 if (G->hasSection()) {
1859 // The kernel uses explicit sections for mostly special global variables
1860 // that we should not instrument. E.g. the kernel may rely on their layout
1861 // without redzones, or remove them at link time ("discard.*"), etc.
1862 if (CompileKernel)
1863 return false;
1864
1865 StringRef Section = G->getSection();
1866
1867 // Globals from llvm.metadata aren't emitted, do not instrument them.
1868 if (Section == "llvm.metadata") return false;
1869 // Do not instrument globals from special LLVM sections.
1870 if (Section.find("__llvm") != StringRef::npos || Section.find("__LLVM") != StringRef::npos) return false;
1871
1872 // Do not instrument function pointers to initialization and termination
1873 // routines: dynamic linker will not properly handle redzones.
1874 if (Section.startswith(".preinit_array") ||
1875 Section.startswith(".init_array") ||
1876 Section.startswith(".fini_array")) {
1877 return false;
1878 }
1879
1880 // Do not instrument user-defined sections (with names resembling
1881 // valid C identifiers)
1882 if (TargetTriple.isOSBinFormatELF()) {
1883 if (llvm::all_of(Section,
1884 [](char c) { return llvm::isAlnum(c) || c == '_'; }))
1885 return false;
1886 }
1887
1888 // On COFF, if the section name contains '$', it is highly likely that the
1889 // user is using section sorting to create an array of globals similar to
1890 // the way initialization callbacks are registered in .init_array and
1891 // .CRT$XCU. The ATL also registers things in .ATL$__[azm]. Adding redzones
1892 // to such globals is counterproductive, because the intent is that they
1893 // will form an array, and out-of-bounds accesses are expected.
1894 // See https://github.com/google/sanitizers/issues/305
1895 // and http://msdn.microsoft.com/en-US/en-en/library/bb918180(v=vs.120).aspx
1896 if (TargetTriple.isOSBinFormatCOFF() && Section.contains('$')) {
1897 LLVM_DEBUG(dbgs() << "Ignoring global in sorted section (contains '$'): "
1898 << *G << "\n");
1899 return false;
1900 }
1901
1902 if (TargetTriple.isOSBinFormatMachO()) {
1903 StringRef ParsedSegment, ParsedSection;
1904 unsigned TAA = 0, StubSize = 0;
1905 bool TAAParsed;
1906 std::string ErrorCode = MCSectionMachO::ParseSectionSpecifier(
1907 Section, ParsedSegment, ParsedSection, TAA, TAAParsed, StubSize);
1908 assert(ErrorCode.empty() && "Invalid section specifier.");
1909
1910 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
1911 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
1912 // them.
1913 if (ParsedSegment == "__OBJC" ||
1914 (ParsedSegment == "__DATA" && ParsedSection.startswith("__objc_"))) {
1915 LLVM_DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G << "\n");
1916 return false;
1917 }
1918 // See https://github.com/google/sanitizers/issues/32
1919 // Constant CFString instances are compiled in the following way:
1920 // -- the string buffer is emitted into
1921 // __TEXT,__cstring,cstring_literals
1922 // -- the constant NSConstantString structure referencing that buffer
1923 // is placed into __DATA,__cfstring
1924 // Therefore there's no point in placing redzones into __DATA,__cfstring.
1925 // Moreover, it causes the linker to crash on OS X 10.7
1926 if (ParsedSegment == "__DATA" && ParsedSection == "__cfstring") {
1927 LLVM_DEBUG(dbgs() << "Ignoring CFString: " << *G << "\n");
1928 return false;
1929 }
1930 // The linker merges the contents of cstring_literals and removes the
1931 // trailing zeroes.
1932 if (ParsedSegment == "__TEXT" && (TAA & MachO::S_CSTRING_LITERALS)) {
1933 LLVM_DEBUG(dbgs() << "Ignoring a cstring literal: " << *G << "\n");
1934 return false;
1935 }
1936 }
1937 }
1938
1939 if (CompileKernel) {
1940 // Globals that prefixed by "__" are special and cannot be padded with a
1941 // redzone.
1942 if (G->getName().startswith("__"))
1943 return false;
1944 }
1945
1946 return true;
1947 }
1948
1949 // On Mach-O platforms, we emit global metadata in a separate section of the
1950 // binary in order to allow the linker to properly dead strip. This is only
1951 // supported on recent versions of ld64.
ShouldUseMachOGlobalsSection() const1952 bool ModuleAddressSanitizer::ShouldUseMachOGlobalsSection() const {
1953 if (!TargetTriple.isOSBinFormatMachO())
1954 return false;
1955
1956 if (TargetTriple.isMacOSX() && !TargetTriple.isMacOSXVersionLT(10, 11))
1957 return true;
1958 if (TargetTriple.isiOS() /* or tvOS */ && !TargetTriple.isOSVersionLT(9))
1959 return true;
1960 if (TargetTriple.isWatchOS() && !TargetTriple.isOSVersionLT(2))
1961 return true;
1962
1963 return false;
1964 }
1965
getGlobalMetadataSection() const1966 StringRef ModuleAddressSanitizer::getGlobalMetadataSection() const {
1967 switch (TargetTriple.getObjectFormat()) {
1968 case Triple::COFF: return ".ASAN$GL";
1969 case Triple::ELF: return "asan_globals";
1970 case Triple::MachO: return "__DATA,__asan_globals,regular";
1971 case Triple::Wasm:
1972 case Triple::GOFF:
1973 case Triple::XCOFF:
1974 report_fatal_error(
1975 "ModuleAddressSanitizer not implemented for object file format");
1976 case Triple::UnknownObjectFormat:
1977 break;
1978 }
1979 llvm_unreachable("unsupported object format");
1980 }
1981
initializeCallbacks(Module & M)1982 void ModuleAddressSanitizer::initializeCallbacks(Module &M) {
1983 IRBuilder<> IRB(*C);
1984
1985 // Declare our poisoning and unpoisoning functions.
1986 AsanPoisonGlobals =
1987 M.getOrInsertFunction(kAsanPoisonGlobalsName, IRB.getVoidTy(), IntptrTy);
1988 AsanUnpoisonGlobals =
1989 M.getOrInsertFunction(kAsanUnpoisonGlobalsName, IRB.getVoidTy());
1990
1991 // Declare functions that register/unregister globals.
1992 AsanRegisterGlobals = M.getOrInsertFunction(
1993 kAsanRegisterGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy);
1994 AsanUnregisterGlobals = M.getOrInsertFunction(
1995 kAsanUnregisterGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy);
1996
1997 // Declare the functions that find globals in a shared object and then invoke
1998 // the (un)register function on them.
1999 AsanRegisterImageGlobals = M.getOrInsertFunction(
2000 kAsanRegisterImageGlobalsName, IRB.getVoidTy(), IntptrTy);
2001 AsanUnregisterImageGlobals = M.getOrInsertFunction(
2002 kAsanUnregisterImageGlobalsName, IRB.getVoidTy(), IntptrTy);
2003
2004 AsanRegisterElfGlobals =
2005 M.getOrInsertFunction(kAsanRegisterElfGlobalsName, IRB.getVoidTy(),
2006 IntptrTy, IntptrTy, IntptrTy);
2007 AsanUnregisterElfGlobals =
2008 M.getOrInsertFunction(kAsanUnregisterElfGlobalsName, IRB.getVoidTy(),
2009 IntptrTy, IntptrTy, IntptrTy);
2010 }
2011
2012 // Put the metadata and the instrumented global in the same group. This ensures
2013 // that the metadata is discarded if the instrumented global is discarded.
SetComdatForGlobalMetadata(GlobalVariable * G,GlobalVariable * Metadata,StringRef InternalSuffix)2014 void ModuleAddressSanitizer::SetComdatForGlobalMetadata(
2015 GlobalVariable *G, GlobalVariable *Metadata, StringRef InternalSuffix) {
2016 Module &M = *G->getParent();
2017 Comdat *C = G->getComdat();
2018 if (!C) {
2019 if (!G->hasName()) {
2020 // If G is unnamed, it must be internal. Give it an artificial name
2021 // so we can put it in a comdat.
2022 assert(G->hasLocalLinkage());
2023 G->setName(Twine(kAsanGenPrefix) + "_anon_global");
2024 }
2025
2026 if (!InternalSuffix.empty() && G->hasLocalLinkage()) {
2027 std::string Name = std::string(G->getName());
2028 Name += InternalSuffix;
2029 C = M.getOrInsertComdat(Name);
2030 } else {
2031 C = M.getOrInsertComdat(G->getName());
2032 }
2033
2034 // Make this IMAGE_COMDAT_SELECT_NODUPLICATES on COFF. Also upgrade private
2035 // linkage to internal linkage so that a symbol table entry is emitted. This
2036 // is necessary in order to create the comdat group.
2037 if (TargetTriple.isOSBinFormatCOFF()) {
2038 C->setSelectionKind(Comdat::NoDuplicates);
2039 if (G->hasPrivateLinkage())
2040 G->setLinkage(GlobalValue::InternalLinkage);
2041 }
2042 G->setComdat(C);
2043 }
2044
2045 assert(G->hasComdat());
2046 Metadata->setComdat(G->getComdat());
2047 }
2048
2049 // Create a separate metadata global and put it in the appropriate ASan
2050 // global registration section.
2051 GlobalVariable *
CreateMetadataGlobal(Module & M,Constant * Initializer,StringRef OriginalName)2052 ModuleAddressSanitizer::CreateMetadataGlobal(Module &M, Constant *Initializer,
2053 StringRef OriginalName) {
2054 auto Linkage = TargetTriple.isOSBinFormatMachO()
2055 ? GlobalVariable::InternalLinkage
2056 : GlobalVariable::PrivateLinkage;
2057 GlobalVariable *Metadata = new GlobalVariable(
2058 M, Initializer->getType(), false, Linkage, Initializer,
2059 Twine("__asan_global_") + GlobalValue::dropLLVMManglingEscape(OriginalName));
2060 Metadata->setSection(getGlobalMetadataSection());
2061 return Metadata;
2062 }
2063
CreateAsanModuleDtor(Module & M)2064 Instruction *ModuleAddressSanitizer::CreateAsanModuleDtor(Module &M) {
2065 AsanDtorFunction =
2066 Function::Create(FunctionType::get(Type::getVoidTy(*C), false),
2067 GlobalValue::InternalLinkage, kAsanModuleDtorName, &M);
2068 BasicBlock *AsanDtorBB = BasicBlock::Create(*C, "", AsanDtorFunction);
2069
2070 return ReturnInst::Create(*C, AsanDtorBB);
2071 }
2072
InstrumentGlobalsCOFF(IRBuilder<> & IRB,Module & M,ArrayRef<GlobalVariable * > ExtendedGlobals,ArrayRef<Constant * > MetadataInitializers)2073 void ModuleAddressSanitizer::InstrumentGlobalsCOFF(
2074 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2075 ArrayRef<Constant *> MetadataInitializers) {
2076 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2077 auto &DL = M.getDataLayout();
2078
2079 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2080 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2081 Constant *Initializer = MetadataInitializers[i];
2082 GlobalVariable *G = ExtendedGlobals[i];
2083 GlobalVariable *Metadata =
2084 CreateMetadataGlobal(M, Initializer, G->getName());
2085 MDNode *MD = MDNode::get(M.getContext(), ValueAsMetadata::get(G));
2086 Metadata->setMetadata(LLVMContext::MD_associated, MD);
2087 MetadataGlobals[i] = Metadata;
2088
2089 // The MSVC linker always inserts padding when linking incrementally. We
2090 // cope with that by aligning each struct to its size, which must be a power
2091 // of two.
2092 unsigned SizeOfGlobalStruct = DL.getTypeAllocSize(Initializer->getType());
2093 assert(isPowerOf2_32(SizeOfGlobalStruct) &&
2094 "global metadata will not be padded appropriately");
2095 Metadata->setAlignment(assumeAligned(SizeOfGlobalStruct));
2096
2097 SetComdatForGlobalMetadata(G, Metadata, "");
2098 }
2099
2100 // Update llvm.compiler.used, adding the new metadata globals. This is
2101 // needed so that during LTO these variables stay alive.
2102 if (!MetadataGlobals.empty())
2103 appendToCompilerUsed(M, MetadataGlobals);
2104 }
2105
InstrumentGlobalsELF(IRBuilder<> & IRB,Module & M,ArrayRef<GlobalVariable * > ExtendedGlobals,ArrayRef<Constant * > MetadataInitializers,const std::string & UniqueModuleId)2106 void ModuleAddressSanitizer::InstrumentGlobalsELF(
2107 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2108 ArrayRef<Constant *> MetadataInitializers,
2109 const std::string &UniqueModuleId) {
2110 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2111
2112 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2113 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2114 GlobalVariable *G = ExtendedGlobals[i];
2115 GlobalVariable *Metadata =
2116 CreateMetadataGlobal(M, MetadataInitializers[i], G->getName());
2117 MDNode *MD = MDNode::get(M.getContext(), ValueAsMetadata::get(G));
2118 Metadata->setMetadata(LLVMContext::MD_associated, MD);
2119 MetadataGlobals[i] = Metadata;
2120
2121 SetComdatForGlobalMetadata(G, Metadata, UniqueModuleId);
2122 }
2123
2124 // Update llvm.compiler.used, adding the new metadata globals. This is
2125 // needed so that during LTO these variables stay alive.
2126 if (!MetadataGlobals.empty())
2127 appendToCompilerUsed(M, MetadataGlobals);
2128
2129 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2130 // to look up the loaded image that contains it. Second, we can store in it
2131 // whether registration has already occurred, to prevent duplicate
2132 // registration.
2133 //
2134 // Common linkage ensures that there is only one global per shared library.
2135 GlobalVariable *RegisteredFlag = new GlobalVariable(
2136 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2137 ConstantInt::get(IntptrTy, 0), kAsanGlobalsRegisteredFlagName);
2138 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2139
2140 // Create start and stop symbols.
2141 GlobalVariable *StartELFMetadata = new GlobalVariable(
2142 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2143 "__start_" + getGlobalMetadataSection());
2144 StartELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2145 GlobalVariable *StopELFMetadata = new GlobalVariable(
2146 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2147 "__stop_" + getGlobalMetadataSection());
2148 StopELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2149
2150 // Create a call to register the globals with the runtime.
2151 IRB.CreateCall(AsanRegisterElfGlobals,
2152 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy),
2153 IRB.CreatePointerCast(StartELFMetadata, IntptrTy),
2154 IRB.CreatePointerCast(StopELFMetadata, IntptrTy)});
2155
2156 // We also need to unregister globals at the end, e.g., when a shared library
2157 // gets closed.
2158 IRBuilder<> IRB_Dtor(CreateAsanModuleDtor(M));
2159 IRB_Dtor.CreateCall(AsanUnregisterElfGlobals,
2160 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy),
2161 IRB.CreatePointerCast(StartELFMetadata, IntptrTy),
2162 IRB.CreatePointerCast(StopELFMetadata, IntptrTy)});
2163 }
2164
InstrumentGlobalsMachO(IRBuilder<> & IRB,Module & M,ArrayRef<GlobalVariable * > ExtendedGlobals,ArrayRef<Constant * > MetadataInitializers)2165 void ModuleAddressSanitizer::InstrumentGlobalsMachO(
2166 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2167 ArrayRef<Constant *> MetadataInitializers) {
2168 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2169
2170 // On recent Mach-O platforms, use a structure which binds the liveness of
2171 // the global variable to the metadata struct. Keep the list of "Liveness" GV
2172 // created to be added to llvm.compiler.used
2173 StructType *LivenessTy = StructType::get(IntptrTy, IntptrTy);
2174 SmallVector<GlobalValue *, 16> LivenessGlobals(ExtendedGlobals.size());
2175
2176 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2177 Constant *Initializer = MetadataInitializers[i];
2178 GlobalVariable *G = ExtendedGlobals[i];
2179 GlobalVariable *Metadata =
2180 CreateMetadataGlobal(M, Initializer, G->getName());
2181
2182 // On recent Mach-O platforms, we emit the global metadata in a way that
2183 // allows the linker to properly strip dead globals.
2184 auto LivenessBinder =
2185 ConstantStruct::get(LivenessTy, Initializer->getAggregateElement(0u),
2186 ConstantExpr::getPointerCast(Metadata, IntptrTy));
2187 GlobalVariable *Liveness = new GlobalVariable(
2188 M, LivenessTy, false, GlobalVariable::InternalLinkage, LivenessBinder,
2189 Twine("__asan_binder_") + G->getName());
2190 Liveness->setSection("__DATA,__asan_liveness,regular,live_support");
2191 LivenessGlobals[i] = Liveness;
2192 }
2193
2194 // Update llvm.compiler.used, adding the new liveness globals. This is
2195 // needed so that during LTO these variables stay alive. The alternative
2196 // would be to have the linker handling the LTO symbols, but libLTO
2197 // current API does not expose access to the section for each symbol.
2198 if (!LivenessGlobals.empty())
2199 appendToCompilerUsed(M, LivenessGlobals);
2200
2201 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2202 // to look up the loaded image that contains it. Second, we can store in it
2203 // whether registration has already occurred, to prevent duplicate
2204 // registration.
2205 //
2206 // common linkage ensures that there is only one global per shared library.
2207 GlobalVariable *RegisteredFlag = new GlobalVariable(
2208 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2209 ConstantInt::get(IntptrTy, 0), kAsanGlobalsRegisteredFlagName);
2210 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2211
2212 IRB.CreateCall(AsanRegisterImageGlobals,
2213 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy)});
2214
2215 // We also need to unregister globals at the end, e.g., when a shared library
2216 // gets closed.
2217 IRBuilder<> IRB_Dtor(CreateAsanModuleDtor(M));
2218 IRB_Dtor.CreateCall(AsanUnregisterImageGlobals,
2219 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy)});
2220 }
2221
InstrumentGlobalsWithMetadataArray(IRBuilder<> & IRB,Module & M,ArrayRef<GlobalVariable * > ExtendedGlobals,ArrayRef<Constant * > MetadataInitializers)2222 void ModuleAddressSanitizer::InstrumentGlobalsWithMetadataArray(
2223 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2224 ArrayRef<Constant *> MetadataInitializers) {
2225 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2226 unsigned N = ExtendedGlobals.size();
2227 assert(N > 0);
2228
2229 // On platforms that don't have a custom metadata section, we emit an array
2230 // of global metadata structures.
2231 ArrayType *ArrayOfGlobalStructTy =
2232 ArrayType::get(MetadataInitializers[0]->getType(), N);
2233 auto AllGlobals = new GlobalVariable(
2234 M, ArrayOfGlobalStructTy, false, GlobalVariable::InternalLinkage,
2235 ConstantArray::get(ArrayOfGlobalStructTy, MetadataInitializers), "");
2236 if (Mapping.Scale > 3)
2237 AllGlobals->setAlignment(Align(1ULL << Mapping.Scale));
2238
2239 IRB.CreateCall(AsanRegisterGlobals,
2240 {IRB.CreatePointerCast(AllGlobals, IntptrTy),
2241 ConstantInt::get(IntptrTy, N)});
2242
2243 // We also need to unregister globals at the end, e.g., when a shared library
2244 // gets closed.
2245 IRBuilder<> IRB_Dtor(CreateAsanModuleDtor(M));
2246 IRB_Dtor.CreateCall(AsanUnregisterGlobals,
2247 {IRB.CreatePointerCast(AllGlobals, IntptrTy),
2248 ConstantInt::get(IntptrTy, N)});
2249 }
2250
2251 // This function replaces all global variables with new variables that have
2252 // trailing redzones. It also creates a function that poisons
2253 // redzones and inserts this function into llvm.global_ctors.
2254 // Sets *CtorComdat to true if the global registration code emitted into the
2255 // asan constructor is comdat-compatible.
InstrumentGlobals(IRBuilder<> & IRB,Module & M,bool * CtorComdat)2256 bool ModuleAddressSanitizer::InstrumentGlobals(IRBuilder<> &IRB, Module &M,
2257 bool *CtorComdat) {
2258 *CtorComdat = false;
2259
2260 // Build set of globals that are aliased by some GA, where
2261 // getExcludedAliasedGlobal(GA) returns the relevant GlobalVariable.
2262 SmallPtrSet<const GlobalVariable *, 16> AliasedGlobalExclusions;
2263 if (CompileKernel) {
2264 for (auto &GA : M.aliases()) {
2265 if (const GlobalVariable *GV = getExcludedAliasedGlobal(GA))
2266 AliasedGlobalExclusions.insert(GV);
2267 }
2268 }
2269
2270 SmallVector<GlobalVariable *, 16> GlobalsToChange;
2271 for (auto &G : M.globals()) {
2272 if (!AliasedGlobalExclusions.count(&G) && shouldInstrumentGlobal(&G))
2273 GlobalsToChange.push_back(&G);
2274 }
2275
2276 size_t n = GlobalsToChange.size();
2277 if (n == 0) {
2278 *CtorComdat = true;
2279 return false;
2280 }
2281
2282 auto &DL = M.getDataLayout();
2283
2284 // A global is described by a structure
2285 // size_t beg;
2286 // size_t size;
2287 // size_t size_with_redzone;
2288 // const char *name;
2289 // const char *module_name;
2290 // size_t has_dynamic_init;
2291 // void *source_location;
2292 // size_t odr_indicator;
2293 // We initialize an array of such structures and pass it to a run-time call.
2294 StructType *GlobalStructTy =
2295 StructType::get(IntptrTy, IntptrTy, IntptrTy, IntptrTy, IntptrTy,
2296 IntptrTy, IntptrTy, IntptrTy);
2297 SmallVector<GlobalVariable *, 16> NewGlobals(n);
2298 SmallVector<Constant *, 16> Initializers(n);
2299
2300 bool HasDynamicallyInitializedGlobals = false;
2301
2302 // We shouldn't merge same module names, as this string serves as unique
2303 // module ID in runtime.
2304 GlobalVariable *ModuleName = createPrivateGlobalForString(
2305 M, M.getModuleIdentifier(), /*AllowMerging*/ false, kAsanGenPrefix);
2306
2307 for (size_t i = 0; i < n; i++) {
2308 GlobalVariable *G = GlobalsToChange[i];
2309
2310 // FIXME: Metadata should be attched directly to the global directly instead
2311 // of being added to llvm.asan.globals.
2312 auto MD = GlobalsMD.get(G);
2313 StringRef NameForGlobal = G->getName();
2314 // Create string holding the global name (use global name from metadata
2315 // if it's available, otherwise just write the name of global variable).
2316 GlobalVariable *Name = createPrivateGlobalForString(
2317 M, MD.Name.empty() ? NameForGlobal : MD.Name,
2318 /*AllowMerging*/ true, kAsanGenPrefix);
2319
2320 Type *Ty = G->getValueType();
2321 const uint64_t SizeInBytes = DL.getTypeAllocSize(Ty);
2322 const uint64_t RightRedzoneSize = getRedzoneSizeForGlobal(SizeInBytes);
2323 Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize);
2324
2325 StructType *NewTy = StructType::get(Ty, RightRedZoneTy);
2326 Constant *NewInitializer = ConstantStruct::get(
2327 NewTy, G->getInitializer(), Constant::getNullValue(RightRedZoneTy));
2328
2329 // Create a new global variable with enough space for a redzone.
2330 GlobalValue::LinkageTypes Linkage = G->getLinkage();
2331 if (G->isConstant() && Linkage == GlobalValue::PrivateLinkage)
2332 Linkage = GlobalValue::InternalLinkage;
2333 GlobalVariable *NewGlobal =
2334 new GlobalVariable(M, NewTy, G->isConstant(), Linkage, NewInitializer,
2335 "", G, G->getThreadLocalMode());
2336 NewGlobal->copyAttributesFrom(G);
2337 NewGlobal->setComdat(G->getComdat());
2338 NewGlobal->setAlignment(MaybeAlign(getMinRedzoneSizeForGlobal()));
2339 // Don't fold globals with redzones. ODR violation detector and redzone
2340 // poisoning implicitly creates a dependence on the global's address, so it
2341 // is no longer valid for it to be marked unnamed_addr.
2342 NewGlobal->setUnnamedAddr(GlobalValue::UnnamedAddr::None);
2343
2344 // Move null-terminated C strings to "__asan_cstring" section on Darwin.
2345 if (TargetTriple.isOSBinFormatMachO() && !G->hasSection() &&
2346 G->isConstant()) {
2347 auto Seq = dyn_cast<ConstantDataSequential>(G->getInitializer());
2348 if (Seq && Seq->isCString())
2349 NewGlobal->setSection("__TEXT,__asan_cstring,regular");
2350 }
2351
2352 // Transfer the debug info and type metadata. The payload starts at offset
2353 // zero so we can copy the metadata over as is.
2354 NewGlobal->copyMetadata(G, 0);
2355
2356 Value *Indices2[2];
2357 Indices2[0] = IRB.getInt32(0);
2358 Indices2[1] = IRB.getInt32(0);
2359
2360 G->replaceAllUsesWith(
2361 ConstantExpr::getGetElementPtr(NewTy, NewGlobal, Indices2, true));
2362 NewGlobal->takeName(G);
2363 G->eraseFromParent();
2364 NewGlobals[i] = NewGlobal;
2365
2366 Constant *SourceLoc;
2367 if (!MD.SourceLoc.empty()) {
2368 auto SourceLocGlobal = createPrivateGlobalForSourceLoc(M, MD.SourceLoc);
2369 SourceLoc = ConstantExpr::getPointerCast(SourceLocGlobal, IntptrTy);
2370 } else {
2371 SourceLoc = ConstantInt::get(IntptrTy, 0);
2372 }
2373
2374 Constant *ODRIndicator = ConstantExpr::getNullValue(IRB.getInt8PtrTy());
2375 GlobalValue *InstrumentedGlobal = NewGlobal;
2376
2377 bool CanUsePrivateAliases =
2378 TargetTriple.isOSBinFormatELF() || TargetTriple.isOSBinFormatMachO() ||
2379 TargetTriple.isOSBinFormatWasm();
2380 if (CanUsePrivateAliases && UsePrivateAlias) {
2381 // Create local alias for NewGlobal to avoid crash on ODR between
2382 // instrumented and non-instrumented libraries.
2383 InstrumentedGlobal =
2384 GlobalAlias::create(GlobalValue::PrivateLinkage, "", NewGlobal);
2385 }
2386
2387 // ODR should not happen for local linkage.
2388 if (NewGlobal->hasLocalLinkage()) {
2389 ODRIndicator = ConstantExpr::getIntToPtr(ConstantInt::get(IntptrTy, -1),
2390 IRB.getInt8PtrTy());
2391 } else if (UseOdrIndicator) {
2392 // With local aliases, we need to provide another externally visible
2393 // symbol __odr_asan_XXX to detect ODR violation.
2394 auto *ODRIndicatorSym =
2395 new GlobalVariable(M, IRB.getInt8Ty(), false, Linkage,
2396 Constant::getNullValue(IRB.getInt8Ty()),
2397 kODRGenPrefix + NameForGlobal, nullptr,
2398 NewGlobal->getThreadLocalMode());
2399
2400 // Set meaningful attributes for indicator symbol.
2401 ODRIndicatorSym->setVisibility(NewGlobal->getVisibility());
2402 ODRIndicatorSym->setDLLStorageClass(NewGlobal->getDLLStorageClass());
2403 ODRIndicatorSym->setAlignment(Align(1));
2404 ODRIndicator = ODRIndicatorSym;
2405 }
2406
2407 Constant *Initializer = ConstantStruct::get(
2408 GlobalStructTy,
2409 ConstantExpr::getPointerCast(InstrumentedGlobal, IntptrTy),
2410 ConstantInt::get(IntptrTy, SizeInBytes),
2411 ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize),
2412 ConstantExpr::getPointerCast(Name, IntptrTy),
2413 ConstantExpr::getPointerCast(ModuleName, IntptrTy),
2414 ConstantInt::get(IntptrTy, MD.IsDynInit), SourceLoc,
2415 ConstantExpr::getPointerCast(ODRIndicator, IntptrTy));
2416
2417 if (ClInitializers && MD.IsDynInit) HasDynamicallyInitializedGlobals = true;
2418
2419 LLVM_DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
2420
2421 Initializers[i] = Initializer;
2422 }
2423
2424 // Add instrumented globals to llvm.compiler.used list to avoid LTO from
2425 // ConstantMerge'ing them.
2426 SmallVector<GlobalValue *, 16> GlobalsToAddToUsedList;
2427 for (size_t i = 0; i < n; i++) {
2428 GlobalVariable *G = NewGlobals[i];
2429 if (G->getName().empty()) continue;
2430 GlobalsToAddToUsedList.push_back(G);
2431 }
2432 appendToCompilerUsed(M, ArrayRef<GlobalValue *>(GlobalsToAddToUsedList));
2433
2434 std::string ELFUniqueModuleId =
2435 (UseGlobalsGC && TargetTriple.isOSBinFormatELF()) ? getUniqueModuleId(&M)
2436 : "";
2437
2438 if (!ELFUniqueModuleId.empty()) {
2439 InstrumentGlobalsELF(IRB, M, NewGlobals, Initializers, ELFUniqueModuleId);
2440 *CtorComdat = true;
2441 } else if (UseGlobalsGC && TargetTriple.isOSBinFormatCOFF()) {
2442 InstrumentGlobalsCOFF(IRB, M, NewGlobals, Initializers);
2443 } else if (UseGlobalsGC && ShouldUseMachOGlobalsSection()) {
2444 InstrumentGlobalsMachO(IRB, M, NewGlobals, Initializers);
2445 } else {
2446 InstrumentGlobalsWithMetadataArray(IRB, M, NewGlobals, Initializers);
2447 }
2448
2449 // Create calls for poisoning before initializers run and unpoisoning after.
2450 if (HasDynamicallyInitializedGlobals)
2451 createInitializerPoisonCalls(M, ModuleName);
2452
2453 LLVM_DEBUG(dbgs() << M);
2454 return true;
2455 }
2456
2457 uint64_t
getRedzoneSizeForGlobal(uint64_t SizeInBytes) const2458 ModuleAddressSanitizer::getRedzoneSizeForGlobal(uint64_t SizeInBytes) const {
2459 constexpr uint64_t kMaxRZ = 1 << 18;
2460 const uint64_t MinRZ = getMinRedzoneSizeForGlobal();
2461
2462 // Calculate RZ, where MinRZ <= RZ <= MaxRZ, and RZ ~ 1/4 * SizeInBytes.
2463 uint64_t RZ =
2464 std::max(MinRZ, std::min(kMaxRZ, (SizeInBytes / MinRZ / 4) * MinRZ));
2465
2466 // Round up to multiple of MinRZ.
2467 if (SizeInBytes % MinRZ)
2468 RZ += MinRZ - (SizeInBytes % MinRZ);
2469 assert((RZ + SizeInBytes) % MinRZ == 0);
2470
2471 return RZ;
2472 }
2473
GetAsanVersion(const Module & M) const2474 int ModuleAddressSanitizer::GetAsanVersion(const Module &M) const {
2475 int LongSize = M.getDataLayout().getPointerSizeInBits();
2476 bool isAndroid = Triple(M.getTargetTriple()).isAndroid();
2477 int Version = 8;
2478 // 32-bit Android is one version ahead because of the switch to dynamic
2479 // shadow.
2480 Version += (LongSize == 32 && isAndroid);
2481 return Version;
2482 }
2483
instrumentModule(Module & M)2484 bool ModuleAddressSanitizer::instrumentModule(Module &M) {
2485 initializeCallbacks(M);
2486
2487 // Create a module constructor. A destructor is created lazily because not all
2488 // platforms, and not all modules need it.
2489 if (CompileKernel) {
2490 // The kernel always builds with its own runtime, and therefore does not
2491 // need the init and version check calls.
2492 AsanCtorFunction = createSanitizerCtor(M, kAsanModuleCtorName);
2493 } else {
2494 std::string AsanVersion = std::to_string(GetAsanVersion(M));
2495 std::string VersionCheckName =
2496 ClInsertVersionCheck ? (kAsanVersionCheckNamePrefix + AsanVersion) : "";
2497 std::tie(AsanCtorFunction, std::ignore) =
2498 createSanitizerCtorAndInitFunctions(M, kAsanModuleCtorName,
2499 kAsanInitName, /*InitArgTypes=*/{},
2500 /*InitArgs=*/{}, VersionCheckName);
2501 }
2502
2503 bool CtorComdat = true;
2504 if (ClGlobals) {
2505 IRBuilder<> IRB(AsanCtorFunction->getEntryBlock().getTerminator());
2506 InstrumentGlobals(IRB, M, &CtorComdat);
2507 }
2508
2509 const uint64_t Priority = GetCtorAndDtorPriority(TargetTriple);
2510
2511 // Put the constructor and destructor in comdat if both
2512 // (1) global instrumentation is not TU-specific
2513 // (2) target is ELF.
2514 if (UseCtorComdat && TargetTriple.isOSBinFormatELF() && CtorComdat) {
2515 AsanCtorFunction->setComdat(M.getOrInsertComdat(kAsanModuleCtorName));
2516 appendToGlobalCtors(M, AsanCtorFunction, Priority, AsanCtorFunction);
2517 if (AsanDtorFunction) {
2518 AsanDtorFunction->setComdat(M.getOrInsertComdat(kAsanModuleDtorName));
2519 appendToGlobalDtors(M, AsanDtorFunction, Priority, AsanDtorFunction);
2520 }
2521 } else {
2522 appendToGlobalCtors(M, AsanCtorFunction, Priority);
2523 if (AsanDtorFunction)
2524 appendToGlobalDtors(M, AsanDtorFunction, Priority);
2525 }
2526
2527 return true;
2528 }
2529
initializeCallbacks(Module & M)2530 void AddressSanitizer::initializeCallbacks(Module &M) {
2531 IRBuilder<> IRB(*C);
2532 // Create __asan_report* callbacks.
2533 // IsWrite, TypeSize and Exp are encoded in the function name.
2534 for (int Exp = 0; Exp < 2; Exp++) {
2535 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
2536 const std::string TypeStr = AccessIsWrite ? "store" : "load";
2537 const std::string ExpStr = Exp ? "exp_" : "";
2538 const std::string EndingStr = Recover ? "_noabort" : "";
2539
2540 SmallVector<Type *, 3> Args2 = {IntptrTy, IntptrTy};
2541 SmallVector<Type *, 2> Args1{1, IntptrTy};
2542 if (Exp) {
2543 Type *ExpType = Type::getInt32Ty(*C);
2544 Args2.push_back(ExpType);
2545 Args1.push_back(ExpType);
2546 }
2547 AsanErrorCallbackSized[AccessIsWrite][Exp] = M.getOrInsertFunction(
2548 kAsanReportErrorTemplate + ExpStr + TypeStr + "_n" + EndingStr,
2549 FunctionType::get(IRB.getVoidTy(), Args2, false));
2550
2551 AsanMemoryAccessCallbackSized[AccessIsWrite][Exp] = M.getOrInsertFunction(
2552 ClMemoryAccessCallbackPrefix + ExpStr + TypeStr + "N" + EndingStr,
2553 FunctionType::get(IRB.getVoidTy(), Args2, false));
2554
2555 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
2556 AccessSizeIndex++) {
2557 const std::string Suffix = TypeStr + itostr(1ULL << AccessSizeIndex);
2558 AsanErrorCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2559 M.getOrInsertFunction(
2560 kAsanReportErrorTemplate + ExpStr + Suffix + EndingStr,
2561 FunctionType::get(IRB.getVoidTy(), Args1, false));
2562
2563 AsanMemoryAccessCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2564 M.getOrInsertFunction(
2565 ClMemoryAccessCallbackPrefix + ExpStr + Suffix + EndingStr,
2566 FunctionType::get(IRB.getVoidTy(), Args1, false));
2567 }
2568 }
2569 }
2570
2571 const std::string MemIntrinCallbackPrefix =
2572 CompileKernel ? std::string("") : ClMemoryAccessCallbackPrefix;
2573 AsanMemmove = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memmove",
2574 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
2575 IRB.getInt8PtrTy(), IntptrTy);
2576 AsanMemcpy = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memcpy",
2577 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
2578 IRB.getInt8PtrTy(), IntptrTy);
2579 AsanMemset = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memset",
2580 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
2581 IRB.getInt32Ty(), IntptrTy);
2582
2583 AsanHandleNoReturnFunc =
2584 M.getOrInsertFunction(kAsanHandleNoReturnName, IRB.getVoidTy());
2585
2586 AsanPtrCmpFunction =
2587 M.getOrInsertFunction(kAsanPtrCmp, IRB.getVoidTy(), IntptrTy, IntptrTy);
2588 AsanPtrSubFunction =
2589 M.getOrInsertFunction(kAsanPtrSub, IRB.getVoidTy(), IntptrTy, IntptrTy);
2590 if (Mapping.InGlobal)
2591 AsanShadowGlobal = M.getOrInsertGlobal("__asan_shadow",
2592 ArrayType::get(IRB.getInt8Ty(), 0));
2593 }
2594
maybeInsertAsanInitAtFunctionEntry(Function & F)2595 bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
2596 // For each NSObject descendant having a +load method, this method is invoked
2597 // by the ObjC runtime before any of the static constructors is called.
2598 // Therefore we need to instrument such methods with a call to __asan_init
2599 // at the beginning in order to initialize our runtime before any access to
2600 // the shadow memory.
2601 // We cannot just ignore these methods, because they may call other
2602 // instrumented functions.
2603 if (F.getName().find(" load]") != std::string::npos) {
2604 FunctionCallee AsanInitFunction =
2605 declareSanitizerInitFunction(*F.getParent(), kAsanInitName, {});
2606 IRBuilder<> IRB(&F.front(), F.front().begin());
2607 IRB.CreateCall(AsanInitFunction, {});
2608 return true;
2609 }
2610 return false;
2611 }
2612
maybeInsertDynamicShadowAtFunctionEntry(Function & F)2613 bool AddressSanitizer::maybeInsertDynamicShadowAtFunctionEntry(Function &F) {
2614 // Generate code only when dynamic addressing is needed.
2615 if (Mapping.Offset != kDynamicShadowSentinel)
2616 return false;
2617
2618 IRBuilder<> IRB(&F.front().front());
2619 if (Mapping.InGlobal) {
2620 if (ClWithIfuncSuppressRemat) {
2621 // An empty inline asm with input reg == output reg.
2622 // An opaque pointer-to-int cast, basically.
2623 InlineAsm *Asm = InlineAsm::get(
2624 FunctionType::get(IntptrTy, {AsanShadowGlobal->getType()}, false),
2625 StringRef(""), StringRef("=r,0"),
2626 /*hasSideEffects=*/false);
2627 LocalDynamicShadow =
2628 IRB.CreateCall(Asm, {AsanShadowGlobal}, ".asan.shadow");
2629 } else {
2630 LocalDynamicShadow =
2631 IRB.CreatePointerCast(AsanShadowGlobal, IntptrTy, ".asan.shadow");
2632 }
2633 } else {
2634 Value *GlobalDynamicAddress = F.getParent()->getOrInsertGlobal(
2635 kAsanShadowMemoryDynamicAddress, IntptrTy);
2636 LocalDynamicShadow = IRB.CreateLoad(IntptrTy, GlobalDynamicAddress);
2637 }
2638 return true;
2639 }
2640
markEscapedLocalAllocas(Function & F)2641 void AddressSanitizer::markEscapedLocalAllocas(Function &F) {
2642 // Find the one possible call to llvm.localescape and pre-mark allocas passed
2643 // to it as uninteresting. This assumes we haven't started processing allocas
2644 // yet. This check is done up front because iterating the use list in
2645 // isInterestingAlloca would be algorithmically slower.
2646 assert(ProcessedAllocas.empty() && "must process localescape before allocas");
2647
2648 // Try to get the declaration of llvm.localescape. If it's not in the module,
2649 // we can exit early.
2650 if (!F.getParent()->getFunction("llvm.localescape")) return;
2651
2652 // Look for a call to llvm.localescape call in the entry block. It can't be in
2653 // any other block.
2654 for (Instruction &I : F.getEntryBlock()) {
2655 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I);
2656 if (II && II->getIntrinsicID() == Intrinsic::localescape) {
2657 // We found a call. Mark all the allocas passed in as uninteresting.
2658 for (Value *Arg : II->arg_operands()) {
2659 AllocaInst *AI = dyn_cast<AllocaInst>(Arg->stripPointerCasts());
2660 assert(AI && AI->isStaticAlloca() &&
2661 "non-static alloca arg to localescape");
2662 ProcessedAllocas[AI] = false;
2663 }
2664 break;
2665 }
2666 }
2667 }
2668
suppressInstrumentationSiteForDebug(int & Instrumented)2669 bool AddressSanitizer::suppressInstrumentationSiteForDebug(int &Instrumented) {
2670 bool ShouldInstrument =
2671 ClDebugMin < 0 || ClDebugMax < 0 ||
2672 (Instrumented >= ClDebugMin && Instrumented <= ClDebugMax);
2673 Instrumented++;
2674 return !ShouldInstrument;
2675 }
2676
instrumentFunction(Function & F,const TargetLibraryInfo * TLI)2677 bool AddressSanitizer::instrumentFunction(Function &F,
2678 const TargetLibraryInfo *TLI) {
2679 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage) return false;
2680 if (!ClDebugFunc.empty() && ClDebugFunc == F.getName()) return false;
2681 if (F.getName().startswith("__asan_")) return false;
2682
2683 bool FunctionModified = false;
2684
2685 // If needed, insert __asan_init before checking for SanitizeAddress attr.
2686 // This function needs to be called even if the function body is not
2687 // instrumented.
2688 if (maybeInsertAsanInitAtFunctionEntry(F))
2689 FunctionModified = true;
2690
2691 // Leave if the function doesn't need instrumentation.
2692 if (!F.hasFnAttribute(Attribute::SanitizeAddress)) return FunctionModified;
2693
2694 LLVM_DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
2695
2696 initializeCallbacks(*F.getParent());
2697
2698 FunctionStateRAII CleanupObj(this);
2699
2700 FunctionModified |= maybeInsertDynamicShadowAtFunctionEntry(F);
2701
2702 // We can't instrument allocas used with llvm.localescape. Only static allocas
2703 // can be passed to that intrinsic.
2704 markEscapedLocalAllocas(F);
2705
2706 // We want to instrument every address only once per basic block (unless there
2707 // are calls between uses).
2708 SmallPtrSet<Value *, 16> TempsToInstrument;
2709 SmallVector<InterestingMemoryOperand, 16> OperandsToInstrument;
2710 SmallVector<MemIntrinsic *, 16> IntrinToInstrument;
2711 SmallVector<Instruction *, 8> NoReturnCalls;
2712 SmallVector<BasicBlock *, 16> AllBlocks;
2713 SmallVector<Instruction *, 16> PointerComparisonsOrSubtracts;
2714 int NumAllocas = 0;
2715
2716 // Fill the set of memory operations to instrument.
2717 for (auto &BB : F) {
2718 AllBlocks.push_back(&BB);
2719 TempsToInstrument.clear();
2720 int NumInsnsPerBB = 0;
2721 for (auto &Inst : BB) {
2722 if (LooksLikeCodeInBug11395(&Inst)) return false;
2723 SmallVector<InterestingMemoryOperand, 1> InterestingOperands;
2724 getInterestingMemoryOperands(&Inst, InterestingOperands);
2725
2726 if (!InterestingOperands.empty()) {
2727 for (auto &Operand : InterestingOperands) {
2728 if (ClOpt && ClOptSameTemp) {
2729 Value *Ptr = Operand.getPtr();
2730 // If we have a mask, skip instrumentation if we've already
2731 // instrumented the full object. But don't add to TempsToInstrument
2732 // because we might get another load/store with a different mask.
2733 if (Operand.MaybeMask) {
2734 if (TempsToInstrument.count(Ptr))
2735 continue; // We've seen this (whole) temp in the current BB.
2736 } else {
2737 if (!TempsToInstrument.insert(Ptr).second)
2738 continue; // We've seen this temp in the current BB.
2739 }
2740 }
2741 OperandsToInstrument.push_back(Operand);
2742 NumInsnsPerBB++;
2743 }
2744 } else if (((ClInvalidPointerPairs || ClInvalidPointerCmp) &&
2745 isInterestingPointerComparison(&Inst)) ||
2746 ((ClInvalidPointerPairs || ClInvalidPointerSub) &&
2747 isInterestingPointerSubtraction(&Inst))) {
2748 PointerComparisonsOrSubtracts.push_back(&Inst);
2749 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&Inst)) {
2750 // ok, take it.
2751 IntrinToInstrument.push_back(MI);
2752 NumInsnsPerBB++;
2753 } else {
2754 if (isa<AllocaInst>(Inst)) NumAllocas++;
2755 if (auto *CB = dyn_cast<CallBase>(&Inst)) {
2756 // A call inside BB.
2757 TempsToInstrument.clear();
2758 if (CB->doesNotReturn() && !CB->hasMetadata("nosanitize"))
2759 NoReturnCalls.push_back(CB);
2760 }
2761 if (CallInst *CI = dyn_cast<CallInst>(&Inst))
2762 maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
2763 }
2764 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB) break;
2765 }
2766 }
2767
2768 bool UseCalls = (ClInstrumentationWithCallsThreshold >= 0 &&
2769 OperandsToInstrument.size() + IntrinToInstrument.size() >
2770 (unsigned)ClInstrumentationWithCallsThreshold);
2771 const DataLayout &DL = F.getParent()->getDataLayout();
2772 ObjectSizeOpts ObjSizeOpts;
2773 ObjSizeOpts.RoundToAlign = true;
2774 ObjectSizeOffsetVisitor ObjSizeVis(DL, TLI, F.getContext(), ObjSizeOpts);
2775
2776 // Instrument.
2777 int NumInstrumented = 0;
2778 for (auto &Operand : OperandsToInstrument) {
2779 if (!suppressInstrumentationSiteForDebug(NumInstrumented))
2780 instrumentMop(ObjSizeVis, Operand, UseCalls,
2781 F.getParent()->getDataLayout());
2782 FunctionModified = true;
2783 }
2784 for (auto Inst : IntrinToInstrument) {
2785 if (!suppressInstrumentationSiteForDebug(NumInstrumented))
2786 instrumentMemIntrinsic(Inst);
2787 FunctionModified = true;
2788 }
2789
2790 FunctionStackPoisoner FSP(F, *this);
2791 bool ChangedStack = FSP.runOnFunction();
2792
2793 // We must unpoison the stack before NoReturn calls (throw, _exit, etc).
2794 // See e.g. https://github.com/google/sanitizers/issues/37
2795 for (auto CI : NoReturnCalls) {
2796 IRBuilder<> IRB(CI);
2797 IRB.CreateCall(AsanHandleNoReturnFunc, {});
2798 }
2799
2800 for (auto Inst : PointerComparisonsOrSubtracts) {
2801 instrumentPointerComparisonOrSubtraction(Inst);
2802 FunctionModified = true;
2803 }
2804
2805 if (ChangedStack || !NoReturnCalls.empty())
2806 FunctionModified = true;
2807
2808 LLVM_DEBUG(dbgs() << "ASAN done instrumenting: " << FunctionModified << " "
2809 << F << "\n");
2810
2811 return FunctionModified;
2812 }
2813
2814 // Workaround for bug 11395: we don't want to instrument stack in functions
2815 // with large assembly blobs (32-bit only), otherwise reg alloc may crash.
2816 // FIXME: remove once the bug 11395 is fixed.
LooksLikeCodeInBug11395(Instruction * I)2817 bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
2818 if (LongSize != 32) return false;
2819 CallInst *CI = dyn_cast<CallInst>(I);
2820 if (!CI || !CI->isInlineAsm()) return false;
2821 if (CI->getNumArgOperands() <= 5) return false;
2822 // We have inline assembly with quite a few arguments.
2823 return true;
2824 }
2825
initializeCallbacks(Module & M)2826 void FunctionStackPoisoner::initializeCallbacks(Module &M) {
2827 IRBuilder<> IRB(*C);
2828 for (int i = 0; i <= kMaxAsanStackMallocSizeClass; i++) {
2829 std::string Suffix = itostr(i);
2830 AsanStackMallocFunc[i] = M.getOrInsertFunction(
2831 kAsanStackMallocNameTemplate + Suffix, IntptrTy, IntptrTy);
2832 AsanStackFreeFunc[i] =
2833 M.getOrInsertFunction(kAsanStackFreeNameTemplate + Suffix,
2834 IRB.getVoidTy(), IntptrTy, IntptrTy);
2835 }
2836 if (ASan.UseAfterScope) {
2837 AsanPoisonStackMemoryFunc = M.getOrInsertFunction(
2838 kAsanPoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy);
2839 AsanUnpoisonStackMemoryFunc = M.getOrInsertFunction(
2840 kAsanUnpoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy);
2841 }
2842
2843 for (size_t Val : {0x00, 0xf1, 0xf2, 0xf3, 0xf5, 0xf8}) {
2844 std::ostringstream Name;
2845 Name << kAsanSetShadowPrefix;
2846 Name << std::setw(2) << std::setfill('0') << std::hex << Val;
2847 AsanSetShadowFunc[Val] =
2848 M.getOrInsertFunction(Name.str(), IRB.getVoidTy(), IntptrTy, IntptrTy);
2849 }
2850
2851 AsanAllocaPoisonFunc = M.getOrInsertFunction(
2852 kAsanAllocaPoison, IRB.getVoidTy(), IntptrTy, IntptrTy);
2853 AsanAllocasUnpoisonFunc = M.getOrInsertFunction(
2854 kAsanAllocasUnpoison, IRB.getVoidTy(), IntptrTy, IntptrTy);
2855 }
2856
copyToShadowInline(ArrayRef<uint8_t> ShadowMask,ArrayRef<uint8_t> ShadowBytes,size_t Begin,size_t End,IRBuilder<> & IRB,Value * ShadowBase)2857 void FunctionStackPoisoner::copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
2858 ArrayRef<uint8_t> ShadowBytes,
2859 size_t Begin, size_t End,
2860 IRBuilder<> &IRB,
2861 Value *ShadowBase) {
2862 if (Begin >= End)
2863 return;
2864
2865 const size_t LargestStoreSizeInBytes =
2866 std::min<size_t>(sizeof(uint64_t), ASan.LongSize / 8);
2867
2868 const bool IsLittleEndian = F.getParent()->getDataLayout().isLittleEndian();
2869
2870 // Poison given range in shadow using larges store size with out leading and
2871 // trailing zeros in ShadowMask. Zeros never change, so they need neither
2872 // poisoning nor up-poisoning. Still we don't mind if some of them get into a
2873 // middle of a store.
2874 for (size_t i = Begin; i < End;) {
2875 if (!ShadowMask[i]) {
2876 assert(!ShadowBytes[i]);
2877 ++i;
2878 continue;
2879 }
2880
2881 size_t StoreSizeInBytes = LargestStoreSizeInBytes;
2882 // Fit store size into the range.
2883 while (StoreSizeInBytes > End - i)
2884 StoreSizeInBytes /= 2;
2885
2886 // Minimize store size by trimming trailing zeros.
2887 for (size_t j = StoreSizeInBytes - 1; j && !ShadowMask[i + j]; --j) {
2888 while (j <= StoreSizeInBytes / 2)
2889 StoreSizeInBytes /= 2;
2890 }
2891
2892 uint64_t Val = 0;
2893 for (size_t j = 0; j < StoreSizeInBytes; j++) {
2894 if (IsLittleEndian)
2895 Val |= (uint64_t)ShadowBytes[i + j] << (8 * j);
2896 else
2897 Val = (Val << 8) | ShadowBytes[i + j];
2898 }
2899
2900 Value *Ptr = IRB.CreateAdd(ShadowBase, ConstantInt::get(IntptrTy, i));
2901 Value *Poison = IRB.getIntN(StoreSizeInBytes * 8, Val);
2902 IRB.CreateAlignedStore(
2903 Poison, IRB.CreateIntToPtr(Ptr, Poison->getType()->getPointerTo()),
2904 Align(1));
2905
2906 i += StoreSizeInBytes;
2907 }
2908 }
2909
copyToShadow(ArrayRef<uint8_t> ShadowMask,ArrayRef<uint8_t> ShadowBytes,IRBuilder<> & IRB,Value * ShadowBase)2910 void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
2911 ArrayRef<uint8_t> ShadowBytes,
2912 IRBuilder<> &IRB, Value *ShadowBase) {
2913 copyToShadow(ShadowMask, ShadowBytes, 0, ShadowMask.size(), IRB, ShadowBase);
2914 }
2915
copyToShadow(ArrayRef<uint8_t> ShadowMask,ArrayRef<uint8_t> ShadowBytes,size_t Begin,size_t End,IRBuilder<> & IRB,Value * ShadowBase)2916 void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
2917 ArrayRef<uint8_t> ShadowBytes,
2918 size_t Begin, size_t End,
2919 IRBuilder<> &IRB, Value *ShadowBase) {
2920 assert(ShadowMask.size() == ShadowBytes.size());
2921 size_t Done = Begin;
2922 for (size_t i = Begin, j = Begin + 1; i < End; i = j++) {
2923 if (!ShadowMask[i]) {
2924 assert(!ShadowBytes[i]);
2925 continue;
2926 }
2927 uint8_t Val = ShadowBytes[i];
2928 if (!AsanSetShadowFunc[Val])
2929 continue;
2930
2931 // Skip same values.
2932 for (; j < End && ShadowMask[j] && Val == ShadowBytes[j]; ++j) {
2933 }
2934
2935 if (j - i >= ClMaxInlinePoisoningSize) {
2936 copyToShadowInline(ShadowMask, ShadowBytes, Done, i, IRB, ShadowBase);
2937 IRB.CreateCall(AsanSetShadowFunc[Val],
2938 {IRB.CreateAdd(ShadowBase, ConstantInt::get(IntptrTy, i)),
2939 ConstantInt::get(IntptrTy, j - i)});
2940 Done = j;
2941 }
2942 }
2943
2944 copyToShadowInline(ShadowMask, ShadowBytes, Done, End, IRB, ShadowBase);
2945 }
2946
2947 // Fake stack allocator (asan_fake_stack.h) has 11 size classes
2948 // for every power of 2 from kMinStackMallocSize to kMaxAsanStackMallocSizeClass
StackMallocSizeClass(uint64_t LocalStackSize)2949 static int StackMallocSizeClass(uint64_t LocalStackSize) {
2950 assert(LocalStackSize <= kMaxStackMallocSize);
2951 uint64_t MaxSize = kMinStackMallocSize;
2952 for (int i = 0;; i++, MaxSize *= 2)
2953 if (LocalStackSize <= MaxSize) return i;
2954 llvm_unreachable("impossible LocalStackSize");
2955 }
2956
copyArgsPassedByValToAllocas()2957 void FunctionStackPoisoner::copyArgsPassedByValToAllocas() {
2958 Instruction *CopyInsertPoint = &F.front().front();
2959 if (CopyInsertPoint == ASan.LocalDynamicShadow) {
2960 // Insert after the dynamic shadow location is determined
2961 CopyInsertPoint = CopyInsertPoint->getNextNode();
2962 assert(CopyInsertPoint);
2963 }
2964 IRBuilder<> IRB(CopyInsertPoint);
2965 const DataLayout &DL = F.getParent()->getDataLayout();
2966 for (Argument &Arg : F.args()) {
2967 if (Arg.hasByValAttr()) {
2968 Type *Ty = Arg.getParamByValType();
2969 const Align Alignment =
2970 DL.getValueOrABITypeAlignment(Arg.getParamAlign(), Ty);
2971
2972 AllocaInst *AI = IRB.CreateAlloca(
2973 Ty, nullptr,
2974 (Arg.hasName() ? Arg.getName() : "Arg" + Twine(Arg.getArgNo())) +
2975 ".byval");
2976 AI->setAlignment(Alignment);
2977 Arg.replaceAllUsesWith(AI);
2978
2979 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2980 IRB.CreateMemCpy(AI, Alignment, &Arg, Alignment, AllocSize);
2981 }
2982 }
2983 }
2984
createPHI(IRBuilder<> & IRB,Value * Cond,Value * ValueIfTrue,Instruction * ThenTerm,Value * ValueIfFalse)2985 PHINode *FunctionStackPoisoner::createPHI(IRBuilder<> &IRB, Value *Cond,
2986 Value *ValueIfTrue,
2987 Instruction *ThenTerm,
2988 Value *ValueIfFalse) {
2989 PHINode *PHI = IRB.CreatePHI(IntptrTy, 2);
2990 BasicBlock *CondBlock = cast<Instruction>(Cond)->getParent();
2991 PHI->addIncoming(ValueIfFalse, CondBlock);
2992 BasicBlock *ThenBlock = ThenTerm->getParent();
2993 PHI->addIncoming(ValueIfTrue, ThenBlock);
2994 return PHI;
2995 }
2996
createAllocaForLayout(IRBuilder<> & IRB,const ASanStackFrameLayout & L,bool Dynamic)2997 Value *FunctionStackPoisoner::createAllocaForLayout(
2998 IRBuilder<> &IRB, const ASanStackFrameLayout &L, bool Dynamic) {
2999 AllocaInst *Alloca;
3000 if (Dynamic) {
3001 Alloca = IRB.CreateAlloca(IRB.getInt8Ty(),
3002 ConstantInt::get(IRB.getInt64Ty(), L.FrameSize),
3003 "MyAlloca");
3004 } else {
3005 Alloca = IRB.CreateAlloca(ArrayType::get(IRB.getInt8Ty(), L.FrameSize),
3006 nullptr, "MyAlloca");
3007 assert(Alloca->isStaticAlloca());
3008 }
3009 assert((ClRealignStack & (ClRealignStack - 1)) == 0);
3010 size_t FrameAlignment = std::max(L.FrameAlignment, (size_t)ClRealignStack);
3011 Alloca->setAlignment(Align(FrameAlignment));
3012 return IRB.CreatePointerCast(Alloca, IntptrTy);
3013 }
3014
createDynamicAllocasInitStorage()3015 void FunctionStackPoisoner::createDynamicAllocasInitStorage() {
3016 BasicBlock &FirstBB = *F.begin();
3017 IRBuilder<> IRB(dyn_cast<Instruction>(FirstBB.begin()));
3018 DynamicAllocaLayout = IRB.CreateAlloca(IntptrTy, nullptr);
3019 IRB.CreateStore(Constant::getNullValue(IntptrTy), DynamicAllocaLayout);
3020 DynamicAllocaLayout->setAlignment(Align(32));
3021 }
3022
processDynamicAllocas()3023 void FunctionStackPoisoner::processDynamicAllocas() {
3024 if (!ClInstrumentDynamicAllocas || DynamicAllocaVec.empty()) {
3025 assert(DynamicAllocaPoisonCallVec.empty());
3026 return;
3027 }
3028
3029 // Insert poison calls for lifetime intrinsics for dynamic allocas.
3030 for (const auto &APC : DynamicAllocaPoisonCallVec) {
3031 assert(APC.InsBefore);
3032 assert(APC.AI);
3033 assert(ASan.isInterestingAlloca(*APC.AI));
3034 assert(!APC.AI->isStaticAlloca());
3035
3036 IRBuilder<> IRB(APC.InsBefore);
3037 poisonAlloca(APC.AI, APC.Size, IRB, APC.DoPoison);
3038 // Dynamic allocas will be unpoisoned unconditionally below in
3039 // unpoisonDynamicAllocas.
3040 // Flag that we need unpoison static allocas.
3041 }
3042
3043 // Handle dynamic allocas.
3044 createDynamicAllocasInitStorage();
3045 for (auto &AI : DynamicAllocaVec)
3046 handleDynamicAllocaCall(AI);
3047 unpoisonDynamicAllocas();
3048 }
3049
3050 /// Collect instructions in the entry block after \p InsBefore which initialize
3051 /// permanent storage for a function argument. These instructions must remain in
3052 /// the entry block so that uninitialized values do not appear in backtraces. An
3053 /// added benefit is that this conserves spill slots. This does not move stores
3054 /// before instrumented / "interesting" allocas.
findStoresToUninstrumentedArgAllocas(AddressSanitizer & ASan,Instruction & InsBefore,SmallVectorImpl<Instruction * > & InitInsts)3055 static void findStoresToUninstrumentedArgAllocas(
3056 AddressSanitizer &ASan, Instruction &InsBefore,
3057 SmallVectorImpl<Instruction *> &InitInsts) {
3058 Instruction *Start = InsBefore.getNextNonDebugInstruction();
3059 for (Instruction *It = Start; It; It = It->getNextNonDebugInstruction()) {
3060 // Argument initialization looks like:
3061 // 1) store <Argument>, <Alloca> OR
3062 // 2) <CastArgument> = cast <Argument> to ...
3063 // store <CastArgument> to <Alloca>
3064 // Do not consider any other kind of instruction.
3065 //
3066 // Note: This covers all known cases, but may not be exhaustive. An
3067 // alternative to pattern-matching stores is to DFS over all Argument uses:
3068 // this might be more general, but is probably much more complicated.
3069 if (isa<AllocaInst>(It) || isa<CastInst>(It))
3070 continue;
3071 if (auto *Store = dyn_cast<StoreInst>(It)) {
3072 // The store destination must be an alloca that isn't interesting for
3073 // ASan to instrument. These are moved up before InsBefore, and they're
3074 // not interesting because allocas for arguments can be mem2reg'd.
3075 auto *Alloca = dyn_cast<AllocaInst>(Store->getPointerOperand());
3076 if (!Alloca || ASan.isInterestingAlloca(*Alloca))
3077 continue;
3078
3079 Value *Val = Store->getValueOperand();
3080 bool IsDirectArgInit = isa<Argument>(Val);
3081 bool IsArgInitViaCast =
3082 isa<CastInst>(Val) &&
3083 isa<Argument>(cast<CastInst>(Val)->getOperand(0)) &&
3084 // Check that the cast appears directly before the store. Otherwise
3085 // moving the cast before InsBefore may break the IR.
3086 Val == It->getPrevNonDebugInstruction();
3087 bool IsArgInit = IsDirectArgInit || IsArgInitViaCast;
3088 if (!IsArgInit)
3089 continue;
3090
3091 if (IsArgInitViaCast)
3092 InitInsts.push_back(cast<Instruction>(Val));
3093 InitInsts.push_back(Store);
3094 continue;
3095 }
3096
3097 // Do not reorder past unknown instructions: argument initialization should
3098 // only involve casts and stores.
3099 return;
3100 }
3101 }
3102
processStaticAllocas()3103 void FunctionStackPoisoner::processStaticAllocas() {
3104 if (AllocaVec.empty()) {
3105 assert(StaticAllocaPoisonCallVec.empty());
3106 return;
3107 }
3108
3109 int StackMallocIdx = -1;
3110 DebugLoc EntryDebugLocation;
3111 if (auto SP = F.getSubprogram())
3112 EntryDebugLocation =
3113 DILocation::get(SP->getContext(), SP->getScopeLine(), 0, SP);
3114
3115 Instruction *InsBefore = AllocaVec[0];
3116 IRBuilder<> IRB(InsBefore);
3117
3118 // Make sure non-instrumented allocas stay in the entry block. Otherwise,
3119 // debug info is broken, because only entry-block allocas are treated as
3120 // regular stack slots.
3121 auto InsBeforeB = InsBefore->getParent();
3122 assert(InsBeforeB == &F.getEntryBlock());
3123 for (auto *AI : StaticAllocasToMoveUp)
3124 if (AI->getParent() == InsBeforeB)
3125 AI->moveBefore(InsBefore);
3126
3127 // Move stores of arguments into entry-block allocas as well. This prevents
3128 // extra stack slots from being generated (to house the argument values until
3129 // they can be stored into the allocas). This also prevents uninitialized
3130 // values from being shown in backtraces.
3131 SmallVector<Instruction *, 8> ArgInitInsts;
3132 findStoresToUninstrumentedArgAllocas(ASan, *InsBefore, ArgInitInsts);
3133 for (Instruction *ArgInitInst : ArgInitInsts)
3134 ArgInitInst->moveBefore(InsBefore);
3135
3136 // If we have a call to llvm.localescape, keep it in the entry block.
3137 if (LocalEscapeCall) LocalEscapeCall->moveBefore(InsBefore);
3138
3139 SmallVector<ASanStackVariableDescription, 16> SVD;
3140 SVD.reserve(AllocaVec.size());
3141 for (AllocaInst *AI : AllocaVec) {
3142 ASanStackVariableDescription D = {AI->getName().data(),
3143 ASan.getAllocaSizeInBytes(*AI),
3144 0,
3145 AI->getAlignment(),
3146 AI,
3147 0,
3148 0};
3149 SVD.push_back(D);
3150 }
3151
3152 // Minimal header size (left redzone) is 4 pointers,
3153 // i.e. 32 bytes on 64-bit platforms and 16 bytes in 32-bit platforms.
3154 size_t Granularity = 1ULL << Mapping.Scale;
3155 size_t MinHeaderSize = std::max((size_t)ASan.LongSize / 2, Granularity);
3156 const ASanStackFrameLayout &L =
3157 ComputeASanStackFrameLayout(SVD, Granularity, MinHeaderSize);
3158
3159 // Build AllocaToSVDMap for ASanStackVariableDescription lookup.
3160 DenseMap<const AllocaInst *, ASanStackVariableDescription *> AllocaToSVDMap;
3161 for (auto &Desc : SVD)
3162 AllocaToSVDMap[Desc.AI] = &Desc;
3163
3164 // Update SVD with information from lifetime intrinsics.
3165 for (const auto &APC : StaticAllocaPoisonCallVec) {
3166 assert(APC.InsBefore);
3167 assert(APC.AI);
3168 assert(ASan.isInterestingAlloca(*APC.AI));
3169 assert(APC.AI->isStaticAlloca());
3170
3171 ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3172 Desc.LifetimeSize = Desc.Size;
3173 if (const DILocation *FnLoc = EntryDebugLocation.get()) {
3174 if (const DILocation *LifetimeLoc = APC.InsBefore->getDebugLoc().get()) {
3175 if (LifetimeLoc->getFile() == FnLoc->getFile())
3176 if (unsigned Line = LifetimeLoc->getLine())
3177 Desc.Line = std::min(Desc.Line ? Desc.Line : Line, Line);
3178 }
3179 }
3180 }
3181
3182 auto DescriptionString = ComputeASanStackFrameDescription(SVD);
3183 LLVM_DEBUG(dbgs() << DescriptionString << " --- " << L.FrameSize << "\n");
3184 uint64_t LocalStackSize = L.FrameSize;
3185 bool DoStackMalloc = ClUseAfterReturn && !ASan.CompileKernel &&
3186 LocalStackSize <= kMaxStackMallocSize;
3187 bool DoDynamicAlloca = ClDynamicAllocaStack;
3188 // Don't do dynamic alloca or stack malloc if:
3189 // 1) There is inline asm: too often it makes assumptions on which registers
3190 // are available.
3191 // 2) There is a returns_twice call (typically setjmp), which is
3192 // optimization-hostile, and doesn't play well with introduced indirect
3193 // register-relative calculation of local variable addresses.
3194 DoDynamicAlloca &= !HasInlineAsm && !HasReturnsTwiceCall;
3195 DoStackMalloc &= !HasInlineAsm && !HasReturnsTwiceCall;
3196
3197 Value *StaticAlloca =
3198 DoDynamicAlloca ? nullptr : createAllocaForLayout(IRB, L, false);
3199
3200 Value *FakeStack;
3201 Value *LocalStackBase;
3202 Value *LocalStackBaseAlloca;
3203 uint8_t DIExprFlags = DIExpression::ApplyOffset;
3204
3205 if (DoStackMalloc) {
3206 LocalStackBaseAlloca =
3207 IRB.CreateAlloca(IntptrTy, nullptr, "asan_local_stack_base");
3208 // void *FakeStack = __asan_option_detect_stack_use_after_return
3209 // ? __asan_stack_malloc_N(LocalStackSize)
3210 // : nullptr;
3211 // void *LocalStackBase = (FakeStack) ? FakeStack : alloca(LocalStackSize);
3212 Constant *OptionDetectUseAfterReturn = F.getParent()->getOrInsertGlobal(
3213 kAsanOptionDetectUseAfterReturn, IRB.getInt32Ty());
3214 Value *UseAfterReturnIsEnabled = IRB.CreateICmpNE(
3215 IRB.CreateLoad(IRB.getInt32Ty(), OptionDetectUseAfterReturn),
3216 Constant::getNullValue(IRB.getInt32Ty()));
3217 Instruction *Term =
3218 SplitBlockAndInsertIfThen(UseAfterReturnIsEnabled, InsBefore, false);
3219 IRBuilder<> IRBIf(Term);
3220 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3221 assert(StackMallocIdx <= kMaxAsanStackMallocSizeClass);
3222 Value *FakeStackValue =
3223 IRBIf.CreateCall(AsanStackMallocFunc[StackMallocIdx],
3224 ConstantInt::get(IntptrTy, LocalStackSize));
3225 IRB.SetInsertPoint(InsBefore);
3226 FakeStack = createPHI(IRB, UseAfterReturnIsEnabled, FakeStackValue, Term,
3227 ConstantInt::get(IntptrTy, 0));
3228
3229 Value *NoFakeStack =
3230 IRB.CreateICmpEQ(FakeStack, Constant::getNullValue(IntptrTy));
3231 Term = SplitBlockAndInsertIfThen(NoFakeStack, InsBefore, false);
3232 IRBIf.SetInsertPoint(Term);
3233 Value *AllocaValue =
3234 DoDynamicAlloca ? createAllocaForLayout(IRBIf, L, true) : StaticAlloca;
3235
3236 IRB.SetInsertPoint(InsBefore);
3237 LocalStackBase = createPHI(IRB, NoFakeStack, AllocaValue, Term, FakeStack);
3238 IRB.CreateStore(LocalStackBase, LocalStackBaseAlloca);
3239 DIExprFlags |= DIExpression::DerefBefore;
3240 } else {
3241 // void *FakeStack = nullptr;
3242 // void *LocalStackBase = alloca(LocalStackSize);
3243 FakeStack = ConstantInt::get(IntptrTy, 0);
3244 LocalStackBase =
3245 DoDynamicAlloca ? createAllocaForLayout(IRB, L, true) : StaticAlloca;
3246 LocalStackBaseAlloca = LocalStackBase;
3247 }
3248
3249 // It shouldn't matter whether we pass an `alloca` or a `ptrtoint` as the
3250 // dbg.declare address opereand, but passing a `ptrtoint` seems to confuse
3251 // later passes and can result in dropped variable coverage in debug info.
3252 Value *LocalStackBaseAllocaPtr =
3253 isa<PtrToIntInst>(LocalStackBaseAlloca)
3254 ? cast<PtrToIntInst>(LocalStackBaseAlloca)->getPointerOperand()
3255 : LocalStackBaseAlloca;
3256 assert(isa<AllocaInst>(LocalStackBaseAllocaPtr) &&
3257 "Variable descriptions relative to ASan stack base will be dropped");
3258
3259 // Replace Alloca instructions with base+offset.
3260 for (const auto &Desc : SVD) {
3261 AllocaInst *AI = Desc.AI;
3262 replaceDbgDeclare(AI, LocalStackBaseAllocaPtr, DIB, DIExprFlags,
3263 Desc.Offset);
3264 Value *NewAllocaPtr = IRB.CreateIntToPtr(
3265 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, Desc.Offset)),
3266 AI->getType());
3267 AI->replaceAllUsesWith(NewAllocaPtr);
3268 }
3269
3270 // The left-most redzone has enough space for at least 4 pointers.
3271 // Write the Magic value to redzone[0].
3272 Value *BasePlus0 = IRB.CreateIntToPtr(LocalStackBase, IntptrPtrTy);
3273 IRB.CreateStore(ConstantInt::get(IntptrTy, kCurrentStackFrameMagic),
3274 BasePlus0);
3275 // Write the frame description constant to redzone[1].
3276 Value *BasePlus1 = IRB.CreateIntToPtr(
3277 IRB.CreateAdd(LocalStackBase,
3278 ConstantInt::get(IntptrTy, ASan.LongSize / 8)),
3279 IntptrPtrTy);
3280 GlobalVariable *StackDescriptionGlobal =
3281 createPrivateGlobalForString(*F.getParent(), DescriptionString,
3282 /*AllowMerging*/ true, kAsanGenPrefix);
3283 Value *Description = IRB.CreatePointerCast(StackDescriptionGlobal, IntptrTy);
3284 IRB.CreateStore(Description, BasePlus1);
3285 // Write the PC to redzone[2].
3286 Value *BasePlus2 = IRB.CreateIntToPtr(
3287 IRB.CreateAdd(LocalStackBase,
3288 ConstantInt::get(IntptrTy, 2 * ASan.LongSize / 8)),
3289 IntptrPtrTy);
3290 IRB.CreateStore(IRB.CreatePointerCast(&F, IntptrTy), BasePlus2);
3291
3292 const auto &ShadowAfterScope = GetShadowBytesAfterScope(SVD, L);
3293
3294 // Poison the stack red zones at the entry.
3295 Value *ShadowBase = ASan.memToShadow(LocalStackBase, IRB);
3296 // As mask we must use most poisoned case: red zones and after scope.
3297 // As bytes we can use either the same or just red zones only.
3298 copyToShadow(ShadowAfterScope, ShadowAfterScope, IRB, ShadowBase);
3299
3300 if (!StaticAllocaPoisonCallVec.empty()) {
3301 const auto &ShadowInScope = GetShadowBytes(SVD, L);
3302
3303 // Poison static allocas near lifetime intrinsics.
3304 for (const auto &APC : StaticAllocaPoisonCallVec) {
3305 const ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3306 assert(Desc.Offset % L.Granularity == 0);
3307 size_t Begin = Desc.Offset / L.Granularity;
3308 size_t End = Begin + (APC.Size + L.Granularity - 1) / L.Granularity;
3309
3310 IRBuilder<> IRB(APC.InsBefore);
3311 copyToShadow(ShadowAfterScope,
3312 APC.DoPoison ? ShadowAfterScope : ShadowInScope, Begin, End,
3313 IRB, ShadowBase);
3314 }
3315 }
3316
3317 SmallVector<uint8_t, 64> ShadowClean(ShadowAfterScope.size(), 0);
3318 SmallVector<uint8_t, 64> ShadowAfterReturn;
3319
3320 // (Un)poison the stack before all ret instructions.
3321 for (Instruction *Ret : RetVec) {
3322 IRBuilder<> IRBRet(Ret);
3323 // Mark the current frame as retired.
3324 IRBRet.CreateStore(ConstantInt::get(IntptrTy, kRetiredStackFrameMagic),
3325 BasePlus0);
3326 if (DoStackMalloc) {
3327 assert(StackMallocIdx >= 0);
3328 // if FakeStack != 0 // LocalStackBase == FakeStack
3329 // // In use-after-return mode, poison the whole stack frame.
3330 // if StackMallocIdx <= 4
3331 // // For small sizes inline the whole thing:
3332 // memset(ShadowBase, kAsanStackAfterReturnMagic, ShadowSize);
3333 // **SavedFlagPtr(FakeStack) = 0
3334 // else
3335 // __asan_stack_free_N(FakeStack, LocalStackSize)
3336 // else
3337 // <This is not a fake stack; unpoison the redzones>
3338 Value *Cmp =
3339 IRBRet.CreateICmpNE(FakeStack, Constant::getNullValue(IntptrTy));
3340 Instruction *ThenTerm, *ElseTerm;
3341 SplitBlockAndInsertIfThenElse(Cmp, Ret, &ThenTerm, &ElseTerm);
3342
3343 IRBuilder<> IRBPoison(ThenTerm);
3344 if (StackMallocIdx <= 4) {
3345 int ClassSize = kMinStackMallocSize << StackMallocIdx;
3346 ShadowAfterReturn.resize(ClassSize / L.Granularity,
3347 kAsanStackUseAfterReturnMagic);
3348 copyToShadow(ShadowAfterReturn, ShadowAfterReturn, IRBPoison,
3349 ShadowBase);
3350 Value *SavedFlagPtrPtr = IRBPoison.CreateAdd(
3351 FakeStack,
3352 ConstantInt::get(IntptrTy, ClassSize - ASan.LongSize / 8));
3353 Value *SavedFlagPtr = IRBPoison.CreateLoad(
3354 IntptrTy, IRBPoison.CreateIntToPtr(SavedFlagPtrPtr, IntptrPtrTy));
3355 IRBPoison.CreateStore(
3356 Constant::getNullValue(IRBPoison.getInt8Ty()),
3357 IRBPoison.CreateIntToPtr(SavedFlagPtr, IRBPoison.getInt8PtrTy()));
3358 } else {
3359 // For larger frames call __asan_stack_free_*.
3360 IRBPoison.CreateCall(
3361 AsanStackFreeFunc[StackMallocIdx],
3362 {FakeStack, ConstantInt::get(IntptrTy, LocalStackSize)});
3363 }
3364
3365 IRBuilder<> IRBElse(ElseTerm);
3366 copyToShadow(ShadowAfterScope, ShadowClean, IRBElse, ShadowBase);
3367 } else {
3368 copyToShadow(ShadowAfterScope, ShadowClean, IRBRet, ShadowBase);
3369 }
3370 }
3371
3372 // We are done. Remove the old unused alloca instructions.
3373 for (auto AI : AllocaVec) AI->eraseFromParent();
3374 }
3375
poisonAlloca(Value * V,uint64_t Size,IRBuilder<> & IRB,bool DoPoison)3376 void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
3377 IRBuilder<> &IRB, bool DoPoison) {
3378 // For now just insert the call to ASan runtime.
3379 Value *AddrArg = IRB.CreatePointerCast(V, IntptrTy);
3380 Value *SizeArg = ConstantInt::get(IntptrTy, Size);
3381 IRB.CreateCall(
3382 DoPoison ? AsanPoisonStackMemoryFunc : AsanUnpoisonStackMemoryFunc,
3383 {AddrArg, SizeArg});
3384 }
3385
3386 // Handling llvm.lifetime intrinsics for a given %alloca:
3387 // (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
3388 // (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
3389 // invalid accesses) and unpoison it for llvm.lifetime.start (the memory
3390 // could be poisoned by previous llvm.lifetime.end instruction, as the
3391 // variable may go in and out of scope several times, e.g. in loops).
3392 // (3) if we poisoned at least one %alloca in a function,
3393 // unpoison the whole stack frame at function exit.
handleDynamicAllocaCall(AllocaInst * AI)3394 void FunctionStackPoisoner::handleDynamicAllocaCall(AllocaInst *AI) {
3395 IRBuilder<> IRB(AI);
3396
3397 const unsigned Alignment = std::max(kAllocaRzSize, AI->getAlignment());
3398 const uint64_t AllocaRedzoneMask = kAllocaRzSize - 1;
3399
3400 Value *Zero = Constant::getNullValue(IntptrTy);
3401 Value *AllocaRzSize = ConstantInt::get(IntptrTy, kAllocaRzSize);
3402 Value *AllocaRzMask = ConstantInt::get(IntptrTy, AllocaRedzoneMask);
3403
3404 // Since we need to extend alloca with additional memory to locate
3405 // redzones, and OldSize is number of allocated blocks with
3406 // ElementSize size, get allocated memory size in bytes by
3407 // OldSize * ElementSize.
3408 const unsigned ElementSize =
3409 F.getParent()->getDataLayout().getTypeAllocSize(AI->getAllocatedType());
3410 Value *OldSize =
3411 IRB.CreateMul(IRB.CreateIntCast(AI->getArraySize(), IntptrTy, false),
3412 ConstantInt::get(IntptrTy, ElementSize));
3413
3414 // PartialSize = OldSize % 32
3415 Value *PartialSize = IRB.CreateAnd(OldSize, AllocaRzMask);
3416
3417 // Misalign = kAllocaRzSize - PartialSize;
3418 Value *Misalign = IRB.CreateSub(AllocaRzSize, PartialSize);
3419
3420 // PartialPadding = Misalign != kAllocaRzSize ? Misalign : 0;
3421 Value *Cond = IRB.CreateICmpNE(Misalign, AllocaRzSize);
3422 Value *PartialPadding = IRB.CreateSelect(Cond, Misalign, Zero);
3423
3424 // AdditionalChunkSize = Alignment + PartialPadding + kAllocaRzSize
3425 // Alignment is added to locate left redzone, PartialPadding for possible
3426 // partial redzone and kAllocaRzSize for right redzone respectively.
3427 Value *AdditionalChunkSize = IRB.CreateAdd(
3428 ConstantInt::get(IntptrTy, Alignment + kAllocaRzSize), PartialPadding);
3429
3430 Value *NewSize = IRB.CreateAdd(OldSize, AdditionalChunkSize);
3431
3432 // Insert new alloca with new NewSize and Alignment params.
3433 AllocaInst *NewAlloca = IRB.CreateAlloca(IRB.getInt8Ty(), NewSize);
3434 NewAlloca->setAlignment(Align(Alignment));
3435
3436 // NewAddress = Address + Alignment
3437 Value *NewAddress = IRB.CreateAdd(IRB.CreatePtrToInt(NewAlloca, IntptrTy),
3438 ConstantInt::get(IntptrTy, Alignment));
3439
3440 // Insert __asan_alloca_poison call for new created alloca.
3441 IRB.CreateCall(AsanAllocaPoisonFunc, {NewAddress, OldSize});
3442
3443 // Store the last alloca's address to DynamicAllocaLayout. We'll need this
3444 // for unpoisoning stuff.
3445 IRB.CreateStore(IRB.CreatePtrToInt(NewAlloca, IntptrTy), DynamicAllocaLayout);
3446
3447 Value *NewAddressPtr = IRB.CreateIntToPtr(NewAddress, AI->getType());
3448
3449 // Replace all uses of AddessReturnedByAlloca with NewAddressPtr.
3450 AI->replaceAllUsesWith(NewAddressPtr);
3451
3452 // We are done. Erase old alloca from parent.
3453 AI->eraseFromParent();
3454 }
3455
3456 // isSafeAccess returns true if Addr is always inbounds with respect to its
3457 // base object. For example, it is a field access or an array access with
3458 // constant inbounds index.
isSafeAccess(ObjectSizeOffsetVisitor & ObjSizeVis,Value * Addr,uint64_t TypeSize) const3459 bool AddressSanitizer::isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis,
3460 Value *Addr, uint64_t TypeSize) const {
3461 SizeOffsetType SizeOffset = ObjSizeVis.compute(Addr);
3462 if (!ObjSizeVis.bothKnown(SizeOffset)) return false;
3463 uint64_t Size = SizeOffset.first.getZExtValue();
3464 int64_t Offset = SizeOffset.second.getSExtValue();
3465 // Three checks are required to ensure safety:
3466 // . Offset >= 0 (since the offset is given from the base ptr)
3467 // . Size >= Offset (unsigned)
3468 // . Size - Offset >= NeededSize (unsigned)
3469 return Offset >= 0 && Size >= uint64_t(Offset) &&
3470 Size - uint64_t(Offset) >= TypeSize / 8;
3471 }
3472